tesis de doct · 2020. 10. 15. · tesis de grado doctorado en economia tapia, stefanoni, jocelyn...
TRANSCRIPT
D O C U M E N T O
D E T R A B A J O
Instituto de EconomíaT
ES
IS d
e D
OC
TO
RA
DO
I N S T I T U T O D E E C O N O M Í A
w w w . e c o n o m i a . p u c . c l
Essays on Knightian Uncertainty
Jocelyn Tapia S.
2019
1
PONTIFICIA UNIVERSIDAD CATOLICAINSTITUTO DE ECONOMIA
DOCTORADO EN ECONOMIA
TESIS DE GRADO
DOCTORADO EN ECONOMIA
Tapia, Stefanoni, Jocelyn Andrea
August, 2019
PONTIFICIA UNIVERSIDAD CATOLICAINSTITUTO DE ECONOMIA
DOCTORADO EN ECONOMIA
Essays on Knightian Uncertainty
Jocelyn Andrea Tapia Stefanoni
Dissertation committee
Klaus Schmidt-Hebbel (advisor), Jaime Casassus (co-advisor), Rodrigo Fuentes (reviewer),
and Martın Uribe (reviewer)
Santiago, August 2019
To my daughters Agustina and Amanda
i
Acknowledgments
I thank the members of my dissertation committee, professors Klaus Schmidt-Hebbel (advisor),
Jaime Casassus (co-advisor), Rodrigo Fuentes (reviewer), and Martin Uribe (reviewer). I thank
my co-authors professors Diederik Aerts and Sandro Sozzo. I am also grateful to professors Sofıa
Bauducco, Rhys Bidder, Jaroslav Borovicka, Jesus Fernandez-Villaverde, Javier Garcıa-Cicco, Lars
Peter Hansen, and Massimo Marinacci for their very helpful comments and discussions. Finally, I
thank Martın Carrasco and Hans Schlechter for their valuable comments.
ii
Contents
1 Introduction 1
2 Welfare Cost of Model Uncertainty in a SOE 4
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Model solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3.1 Second-order approximation of the value function . . . . . . . . . . . . . . . . 12
2.3.2 Welfare analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3.3 Welfare cost calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4 Model calibration and numerical calculations . . . . . . . . . . . . . . . . . . . . . . 15
2.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3 Identifying Quantum Structures in the Ellsberg Paradox 24
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2 A quantum model in Hilbert space for the Ellsberg paradox . . . . . . . . . . . . . . 27
3.2.1 The conceptual Ellsberg entity . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2.2 Modeling acts and utility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2.3 Decision making and superposition states . . . . . . . . . . . . . . . . . . . . 30
3.3 An experiment testing the Ellsberg paradox . . . . . . . . . . . . . . . . . . . . . . . 33
3.4 Quantum modeling the experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.5 A real vector space analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
A Appendix Welfare Cost of Model Uncertainty in a SOE 44
A.1 The model’s first-order conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
A.1.1 Model’s deterministic steady state solution . . . . . . . . . . . . . . . . . . . 44
A.2 Model solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
A.3 Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
iii
List of Figures
2.1 Welfare loss estimations due to aggregate uncertainty, for different parameter val-
ues of risk aversion and concern for model uncertainty; no financial frictions case
(Percentage loss in long-term mean consumption, τ) . . . . . . . . . . . . . . . . . . 17
2.2 Welfare loss estimations due to aggregate uncertainty, for different parameter values
of risk aversion and concern for model uncertainty; financial frictions case (Percentage
loss in long-term mean consumption, τ) . . . . . . . . . . . . . . . . . . . . . . . . . 18
iv
List of Tables
2.1 Calibration of model parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Welfare loss estimations due to aggregate uncertainty, for different parameter values
of risk aversion and concern for model uncertainty (Percentage loss in long-term
mean consumption, τ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.1 The payoff matrix for the Ellsberg paradox situation. . . . . . . . . . . . . . . . . . . 25
A.1 Welfare loss estimations due to aggregate uncertainty, for different combinations of
risk aversion and concern for model uncertainty; no financial frictions case (consump-
tion equivalent variation, τ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
A.2 Welfare loss estimations due to aggregate uncertainty, for different combination of
risk aversion and concern for model uncertainty; with financial frictions (consumption
equivalent variation, τ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
v
Chapter 1
Introduction
Frank Knight (1921) analyzed the problem of an investor, establishing - for the first time in
economics - the key difference between risk and uncertainty. He defined the term “risk” to describe
situations with stochastic outcomes that are described by known probabilities, while using the
term “uncertainty” to refer to situations with outcomes described by unknown probabilities. The
denomination “Knightian uncertainty” comes from this distinction. After Knight (1921), Savage
(1954) contributed an axiomatic treatment of decision making in which preferences over gambles
could be represented by maximizing expected utility under subjective probabilities. Savage’s work
extended the earlier justification of expected utility by von Neumann and Morgenstern (1944)
that had assumed known objective probabilities. Savage’s axioms justify subjective assignments of
probabilities when they are not known in advance (Hansen and Sargent, 2010).
Daniel Ellsberg (1961) presents his famous experiment showing empirically the existence of
ambiguity aversion, when agents prefer to face known probabilities rather than unknown or vague
probabilities, which is not compatible with Savage’s model. Daniel Ellsberg refers to the term
ambiguity instead of Knightian uncertainty following Arrow (1951), who distinguishes between two
sources of uncertainty: (i) “Risk within a model”, where uncertainty is about the outcomes that
emerge in accordance to a probability model that specifies fully the outcome probabilities, and (ii)
“Ambiguity among models”, where uncertainty is about which alternative model should be used to
assign probabilities. Recently, Hansen and Marinacci (2016) have extended the distinction presented
by Arrow adding a third type of uncertainty named as“model misspecification”(also known as model
uncertainty). This uncertainty arises when the true model is not assumed to be among the original
set of models under consideration; then uncertainty is induced by the approximate nature of the
models under consideration used in assigning probabilities.
This thesis is comprised by two essays related to decision-making under uncertainty in the
sense of Frank Knight (1921). In a nutshell, the first essay introduces model uncertainty in an
otherwise standard small open economy real business cycle model, and the second essay presents
an alternative model of decision-making under ambiguity characterizing uncertainty by quantum
probabilities.
The first essay extends the canonical small open-economy model of Schmitt-Grohe and Uribe
1
CHAPTER 1. INTRODUCTION 2
(2003), which is an expected-utility model for classical uncertainty (or risk), to non-expected util-
ity considering model uncertainty. Domestic households have multiplier preferences (Hansen and
Sargent 2008), which leads them to take robust decisions in response to possible model misspeci-
fication for the economy’s aggregate productivity. Using perturbation methods, this essay extends
the literature on RBC models by deriving a closed-form solution for the combined welfare effect of
the two sources of uncertainty (risk and model uncertainty). While classical risk has an ambiguous
effect on welfare (as shown by Cho et al. 2015), the addition of model uncertainty is unambiguously
welfare-deteriorating. Hence the overall effect of uncertainty on welfare is ambiguous, depending on
consumer preferences and model parameters. The paper provides numerical results for the welfare
effects of uncertainty measured by units of consumption equivalence. At moderate (high) levels of
risk aversion, the effect of risk on household welfare is positive (negative). The addition of model
uncertainty - for all levels of concern about model uncertainty and most risk aversion values - turns
the overall effect of uncertainty on household welfare negative. My first result for numerical sim-
ulations shows that for slightly risk-averse households, productivity shocks are in general welfare
improving. However, if they are intensely risk averse, the welfare cost of productivity risk is esti-
mated at 0.007% of long-term consumption, similar to the value of 0.008% found by Lucas (1987).
The second finding is that by introducing model uncertainty, even when the agent is only mildly
risk averse, the overall welfare cost of uncertainty becomes welfare-deteriorating. The third result,
based on numerical calculations, shows that the total effect of uncertainty on welfare depends on
the interaction of risk and model uncertainty. Welfare costs are increasing in both risk aversion
and the concern for model uncertainty. Considering a moderately high degree of risk aversion and
a reasonable concern for model uncertainty, the overall effect of uncertainty on welfare amounts to
a loss of 0.4% of long-term consumption, which is two orders of magnitude higher than the finding
by Lucas. If I consider the case of an agent with very high concern for model misspecification, the
welfare cost reaches a staggering 2.3% of long-term consumption. Finally, I consider the existence
of financial frictions, reflected in a high debt sensitivity of the sovereign risk premium, which has
been proven to increase the welfare cost of productivity shocks in a small open economy, for the
case of classical risk aversion (Ericson and Liu 2012). In a risk aversion context, the largest simu-
lated welfare cost of business cycles is 0.013%, which is significantly smaller than the largest value
obtained when including preferences for robustness. Therefore, with or without financial frictions,
households with a preference for robustness are willing to sacrifice a lot to live in an economy
without fear of misspecification. It is important to remark that the analytical decomposition and
combination of the effects of the two types of uncertainty considered here and the resulting effect
on overall welfare have not been derived in the previous literature on small open economies.
The second essay, co-authored with professors Diederik Aerts and Sandro Sozzo1, elaborates
a quantum model to describe the Ellsberg paradox, showing that the mathematical formalism of
quantum mechanics is capable to represent the ambiguity present in this kind of situations, in terms
of the appearance of typical quantum effects as interference and superposition. Then we analyze
1Aerts, Sozzo and Tapia (2014): Identifying Quantum Structures in the Ellsberg Paradox. International Journalof Theoretical Physics, Vol. 53(10), pp 3666-3682.
CHAPTER 1. INTRODUCTION 3
the data collected in a concrete experiment performed on the Ellsberg paradox and work out a
representation of them in a complex Hilbert space. We are able to represent elections in Ellsberg’s
urn by considering superposition of priors, even when each individual prior cannot explain the
empirically observed behavior. In our model, even when agents are neutral to risk, they can be
ambiguity-averse.
Chapter 2
Welfare Cost of Model Uncertainty in
a Small Open Economy
Abstract
This paper extends the canonical small open-economy model of Schmitt-Grohe and Uribe (2003),
which is an expected-utility model for classical uncertainty (or risk), to non-expected utility con-
sidering model uncertainty. Domestic households have multiplier preferences (Hansen and Sargent
2008), which leads them to take robust decisions in response to possible model misspecification
for the economy’s aggregate productivity. Using perturbation methods, the paper extends the lit-
erature on RBC models by deriving a closed-form solution for the combined welfare effect of the
two sources of uncertainty (risk and model uncertainty). While classical risk has an ambiguous ef-
fect on welfare (as shown by Cho et al. 2015), the addition of model uncertainty is unambiguously
welfare-deteriorating. Hence the overall effect of uncertainty on welfare is ambiguous, depending on
consumer preferences and model parameters. The paper provides numerical results for the welfare
effects of uncertainty measured by units of consumption equivalence. At moderate (high) levels of
risk aversion, the effect of risk on household welfare is positive (negative). The addition of model
uncertainty − for all levels of concern about model uncertainty and most risk aversion values −turns the overall effect of uncertainty on household welfare negative. It is important to remark
that the analytical decomposition and combination of the effects of the two types of uncertainty
considered here and the resulting ambiguous effect on overall welfare have not been derived in the
previous literature on small open economies.
2.1 Introduction
There is a long tradition in assessing the welfare cost of consumption variability, due to dif-
ferent sources of uncertainty. Lucas (1987) started this research area, considering a representative
agent who obtains utility only from consumption, assuming expected utility, constant relative risk
aversion, and a trend-stationary stochastic process of the log of consumption. Lucas finds a small
4
CHAPTER 2. WELFARE COST OF MODEL UNCERTAINTY IN A SOE 5
effect of uncertainty on welfare: the percentage increase in the mean of consumption required to
leave consumers indifferent between consumption fluctuations and a perfectly smooth consumption
path is 0.008%1.
This paper contributes to the literature on the welfare cost of business-cycle fluctuations in a
small open economy (SOE) by introducing model uncertainty. Households face two types of uncer-
tainty: classical uncertainty about the state of future aggregate productivity2 and model uncertainty
about the probability model that describes the movements of productivity. The characterization of
the small open economy follows Schmitt-Grohe and Uribe (2003), augmented by multiplier prefer-
ences (Hansen and Sargent 2008). These preferences correspond to a non-expected utility approach
that reflects a concern for model misspecification (model uncertainty). Households are endowed
with a reference probability model about the stochastic variable (aggregate productivity), but they
distrust its accuracy, considering the possibility that the model is misspecified in a way that is
difficult to detect statistically. Consequently, households want to take decisions that are robust to
model misspecification, by solving a max-min problem. Households maximize utility with respect
to consumption, labor supply, and investment. Households face a malevolent nature that minimizes
the expected utility choosing an unfavorable probabilistic distribution, called the worst case distri-
bution. Nevertheless, the malevolent nature is penalized for the deviations from the reference model
by means of conditional relative entropy between the reference and the distorted models. The solu-
tion of the minimization leads to a recursive value function, similar to Epstein and Zin (1989) and
Weil (1990) preferences (succinctly EZW preferences) and the risk adjustment in the continuation
value for risk sensitive preferences. This paper decomposes the effects of doubts and volatility on
total welfare, deriving a closed-form solution for the cost due to stochastic consumption fluctuations
and the cost derived from uncertainty about the stochastic representation of productivity shocks.
While consumption fluctuations (volatility) have an ambiguous effect on welfare (as shown by Cho
et al., 2015), the additional model uncertainty is unambiguously welfare-deteriorating. Hence the
overall effect of uncertainty on welfare is ambiguous, depending on consumer preferences and model
parameters. Then, the paper provides numerical results for the welfare effects of uncertainty. The
simulations show that when classical risk has a positive effect on welfare (due to a low risk aversion),
addition of model uncertainty can reverse the sign of the overall welfare effect. The total effect of
uncertainty, including both sources, is shown to be negative for a wide range of parameter values.
The relevance of this paper’s results lies in the broad discussion about macroeconomic stabiliza-
tion, as highlighted by Lucas (1987, 2003). If the welfare cost of business cycles is negligible or if
households prefer economic uncertainty, there is no space for counter-cyclical policy in small open
economies. My first result shows that for slightly risk-averse households, productivity shocks are in
general welfare improving. However, if they are intensely risk averse, the welfare cost of productiv-
ity risk is estimated at 0.007% of long-term consumption, similar to the value of 0.008% found by
1Lucas finds that the cost of consumption variability is proportional to risk aversion and to the variance of shocks.His 0.008% estimate corresponds to unitary risk aversion and a shock variance equal to 0.013.
2Garcıa-Cicco, Pancrazi, and Uribe (2010) identify productivity as the main driving factor of business cycles insmall open economies.
CHAPTER 2. WELFARE COST OF MODEL UNCERTAINTY IN A SOE 6
Lucas (1987). The second finding is that by introducing model uncertainty, even when the agent is
only mildly risk averse, the overall welfare cost of uncertainty becomes welfare-deteriorating. My
third result, based on numerical calculations, shows that the total effect of uncertainty on welfare
depends on the interaction of risk and model uncertainty. Welfare costs are increasing in both
risk aversion and the concern for model uncertainty. Considering a moderately high degree of risk
aversion and a reasonable concern for model uncertainty, the overall effect of uncertainty on welfare
amounts to a loss of 0.4% of long-term consumption, which is two orders of magnitude higher than
the finding by Lucas. If I consider the case of an agent with very high concern for model misspec-
ification, the welfare cost reaches a staggering 2.3% of long-term consumption. Finally, I consider
the existence of financial frictions, reflected in a high debt sensitivity of the sovereign risk premium,
which has been proven to increase the welfare cost of productivity shocks in a SOE, for the case of
classical risk aversion (Ericson and Liu 2012). In a risk aversion context the largest simulated wel-
fare cost of business cycles is 0.013%3, which is significantly smaller than the largest value obtained
when including preferences for robustness. Therefore, in both cases households with a preference
for robustness are willing to sacrifice a lot to live in an economy without fear of misspecification.
I discuss the literature related to my research next. Since Lucas’ seminal work, the small cost
of consumption fluctuations estimated by him has motivated many authors to extend his model in
different ways. One branch of these extensions assumes traditional household preferences charac-
terized by expected utility and time separability. Imrohoroglu (1989), Krusell and Smith (1999),
and Krusell et al. (2009) extend Lucas’ model by including uninsurable individual risk and incom-
plete markets, finding larger welfare costs of consumption variability. Alvarez and Jermann (2004)
consider fluctuations in asset prices reporting a small cost of variability on welfare. Subsequent
research shows that consumption volatility is not always welfare deteriorating when output is en-
dogenous and there are multiplicative productivity shocks. Cho et al. (2015) identify two effects
of productivity risk on welfare in production economies: the fluctuation effect and the mean effect.
Whereas the first effect is always detrimental to welfare of risk averse agents, the second may in-
crease welfare. When households respond to larger productivity uncertainty by working harder and
investing more, the mean values of output and consumption could increase with higher uncertainty.
Hence the final effect of uncertainty on welfare is ambiguous. In an open economy, where adjusting
capital is easier than in the closed economy, the positive mean effect is larger than in a closed
economy. In the same vein, Lester et al. (2014) show that welfare can be increasing in volatility of
exogenous shocks in production economies where factor supply is sufficiently elastic.
Closely related to this investigation, a branch of literature specifies recursive utility without
time-separable preferences. In this context, there are three main modeling approaches: EZW
preferences, risk sensitivity, and robust control. Obstfeld (1994), Dolmas (1998), and Epaulard
and Pommeret (2003) use EZW preferences to separate between the effect of risk aversion and the
intertemporal elasticity of substitution to derive the welfare cost of business-cycle fluctuations. In
a nutshell, they underscore the importance of shock persistence and intertemporal substitutability:
the welfare cost of uncertainty increases with both risk aversion and the intertemporal elasticity
3Table 4 in Appendix C presents welfare calculations for a broad set of parameters.
CHAPTER 2. WELFARE COST OF MODEL UNCERTAINTY IN A SOE 7
of substitution. Xu (2017) examines the effects of stochastic volatility on welfare for a continuous-
time extension of EZW preferences and shows that volatility risk can improve welfare, depending
on model parameters.
Tallarini (2000) introduces risk sensitivity with the aim to match financial asset prices, while not
affecting his model’s ability to account for aggregate fluctuations. Assuming a unitary intertemporal
elasticity of substitution4, Tallarini shows that the risk-free rate and the market price of risk are
better matched with augmented risk aversion, and this does not make consumption smoother.
Moreover, the welfare cost of business-cycle fluctuations increases when preference parameters are
chosen to match financial data, and the welfare implications are larger for the production economy
than for the endowment economy.
In general, the advantage of introducing recursive utility as EZW preferences and risk sensitivity
is that it allows to match models with financial data. Expected utility models with high risk aversion
deliver a high market price of risk (addressing the equity premium puzzle), as observed in macro-
financial data, but also raise the risk-free interest rate (not addressing the risk-free rate puzzle). On
the other hand, recursive utility can reconcile the high market price of risk with reasonable values
of the risk-free rate, because it allows to separate between risk aversion and the intertemporal
elasticity of substitution. However, the value for the risk aversion parameter for explaining the
behavior of consumption and asset prices observed in the data is very high. Lucas (2003) discusses
the latter findings, stating that no one has found risk aversion parameters of the required size to
match the data, and concludes: “It would be good to have the equity premium puzzle resolved, but
I think we need to look beyond high estimates of risk aversion to do it.”
Barillas et al. (2009) determine the welfare benefits of removing model uncertainty for a closed
economy, reinterpreting part of the market price of risk as a market price of model uncertainty.
Using the model of error detection to calibrate the concern for model misspecification, they show
that modest amounts of model uncertainty can substitute for large amounts of risk aversion in terms
of choices and asset prices. Therefore this approach allows to reconcile the high market price of risk
with reasonable values of the risk-free rate without resorting to high risk aversion, and the welfare
cost of business cycle calculation is as big as that presented by Tallarini (2000). In the same vein,
Ellison and Sargent (2015) asses the combined effect of idiosyncratic risk and robustness. They find
that individual risk has a larger impact on the cost of business cycles if agents have preferences for
robustness, and the combined effect exceeds the sum of individual effects. As opposed to Barillas
et al. (2009) and Ellison and Sargent (2015), who consider a closed economy and an exogenous
stochastic process for consumption, this paper focuses on a small open production economy with
optimal endogenous consumption.
Regarding solution of the model, this paper is related to the literature on perturbation methods
for recursive utility. I perform a second-order approximation of the value function and equilibrium
conditions (which depend on the value function). Up to a first-order approximation, the equilib-
rium conditions produced by recursive utility or expected utility are strictly equivalent, generating
4This hypothesis allows Tallarini to solve the production economy model by approximation, using the linear-quadratic method developed by Hansen et al. (1999).
CHAPTER 2. WELFARE COST OF MODEL UNCERTAINTY IN A SOE 8
what Levin et al. (2008) define as macroeconomic equivalence. But the underlying microeconomic
differences between first-order equivalent models become important when optimal policy is derived
(microeconomic dissonance). Furthermore, according to Schmitt-Grohe and Uribe (2004), first-
order techniques are not well suited to handle welfare comparisons across alternative stochastic or
policy environments, even without recursive preferences. Caldara et al. (2012) compare different
solution methods for computing equilibria of dynamic stochastic general equilibrium (DSGE) mod-
els with recursive preferences. Their main finding is that perturbation methods are competitive
with projection methods and value function iterations, in terms of accuracy, while being several
orders of magnitude faster to run.
The paper is organized as follows. Section 2 presents the theoretical RBC model for the small
open economy, describing in detail agent preferences and the characterization of the agent’s alterna-
tive probabilistic models. Section 3 describes the solution method, derives the model solution, and
presents the analytical welfare analysis. Section 4 performs numerical calculations for the welfare
cost of classical and model uncertainty. Section 5 concludes.
2.2 The model
The model extends on the canonical Small-Open-Economy Real-Business-Cycle (SOE-RBC),
model introduced by Schmitt-Grohe and Uribe (2003). In the SOE-RBC model, productivity
shocks drive business cycles in a single-good and single-asset production economy. Here, the SOE-
RBC model is extended by considering household preferences for robustness to possible model
misspecification.
Extending the distinction put forward by Arrow (1951), Hansen and Marinacci (2016) distin-
guish between three sources of uncertainty: “(i) Risk within a model, where uncertainty is about
the outcomes that emerge in accordance to a probability model that specifies fully the outcome
probabilities. (ii) Ambiguity among models, where uncertainty is about which alternative model
should be used to assign the probabilities. (iii) Model misspecification, where uncertainty is induced
by the approximate nature of the models under consideration used in assigning probabilities.”
This research is related to the third source of uncertainty, which is termed model uncertainty.
Households have multiplier preferences (Hansen and Sargent 2001, 2008, 2010)5 regarding produc-
tivity (represented by a Hicks neutral technical process): they have a benchmark model for this
stochastic variable but they do not fully trust it. Households acknowledge that the benchmark
model is an approximation of the true data-generating process and could be misspecified in some
way. This leads them to take robust decisions that perform well across a variety of probability
models “near” the benchmark. Agents locate their approximating model within a set of alternative
models that are statistically nearby and which are probabilistic models characterized in terms of the
distortions from the benchmark model. The distortions can be represented in terms of martingales
that twist the benchmark measure in order to obtain absolutely continuous measures that represent
5Multiplier preferences are a particular case of variational preferences, as discussed by Maccheroni et al. (2006a,2006b).
CHAPTER 2. WELFARE COST OF MODEL UNCERTAINTY IN A SOE 9
the alternative models.
Let Ft be the information available at time t, π(εt+1) the benchmark density of shocks εt+1 and
π(εt+1 | Ft) an admissible alternative density conditioned on available information. The likelihood
ratio between the alternative density of shocks and the benchmark probabilistic model is mt+1 =π(εt+1|Ft)π(εt+1) , is non-negative and E(mt+1 | Ft) = 16. The characterization of alternative models
using the likelihood ratio allows to calculate the distorted expectation of a stochastic variable
Wt+1 in period t as E[Wt+1 | Ft] = Et[Wt+1] = Et[mt+1Wt+1]7. The conditional relative entropy
E[mt logmt | Ft], known in econometrics as the Kullback-Leibler divergence, measures discrepancies
between the alternative and reference models.
Multiplier preferences, represented by the Bellman equation (2.1), are defined in terms of a
parameter θ that penalizes discrepancies between the distorted and reference models. These pref-
erences present a max-min problem, where the households maximize utility choosing consumption,
labor supply, and investment, while an evil agent minimizes utility by his choice of the worst prob-
abilistic scenario. The max-min optimization is subject to the agents’ flow budget constraint (2.2),
technology (2.3), the law of motion of capital (2.4), the stochastic process for aggregate productivity
(2.6), and a restriction that imposes a unitary expected value of the likelihood ratio between the
reference and distorted probability measures (2.7).
V (dt, kt, at) = maxct,ht,kt+1,dt+1
minmt+1
u(ct, ht) + βEt[mt+1V (dt+1, kt+1, at+1) + θmt+1 logmt+1] (2.1)
s.t. dt+1 = (1 + rt)dt − yt + ct + it +φ
2(kt+1 − kt)2 (2.2)
yt = atkαt h
1−αt (2.3)
kt+1 = (1− δ)kt + it (2.4)
rt = r∗ + ψ(edt−d − 1) (2.5)
at+1 = (1− ρ)ass + ρat + ηεt+1 (2.6)
Et[mt+1] = 1 (2.7)
where ct represents consumption, ht denotes labor supply, β is the subjective discount factor, dt
denotes the households’ debt position, rt represents the interest rate at which domestic residents can
borrow abroad, yt denotes domestic output, it represents gross investment, kt is physical capital,
and u(ct, ht) is the concave period utility function, which is increasing in ct and decreasing in ht.
The term φ2 (kt+1 − kt)2 in equation (2.2) captures capital adjustment costs that avoid exces-
sive investment volatility in response to shocks in productivity or in the foreign interest rate. The
production function (2.3) is a Cobb-Douglas function, implying a unitary elasticity of substitu-
tion between labor and capital. The stock of physical capital evolves according to (2.4), where δ
represents the rate of depreciation of capital.
The domestic interest rate rt define in equation (2.5) is imperfectly arbitraged to the world
6E(mt+1 | Ft) =∫ π(εt+1|Ft)
π(εt+1)π(εt+1)dεt+1 = 1, integrated with respect to the Lebesgue measure.
7Similar to the change of measure for the transformation to risk-neutral probabilities used in asset pricing.
CHAPTER 2. WELFARE COST OF MODEL UNCERTAINTY IN A SOE 10
interest rate r∗, with a country interest rate premium that is debt elastic and takes the form
ψ(edt−d − 1)8. Here dt is the domestic cross-sectional average level of debt which is considered
exogenous by households, d represents the steady-state level of average debt and ψ > 0 denotes the
sensitivity of the risk premium to deviations of average debt from its steady-state value.
The law of motion of productivity is given by a first-order autoregressive process (equation
(2.6)), where εt is the stochastic productivity shock, which is normally distributed with zero mean
and unit standard deviation. The parameter η defines the standard deviation of the innovation, ass
is the steady state value of productivity, and coefficient ρ reflects the persistence of shocks. The
exogenous process for productivity is specified in levels and not in logs (in contrast to the original
SOE-RBC model by Schmitt-Grohe and Uribe), in order to focus exclusively on the implications
of volatility for welfare9. It is also possible to use logs, as Cho et al. (2015), correcting the process
mean in order to obtain a mean-preserving spread10, but this procedure introduces right skewness
to the distribution of productivity, which tends to be welfare increasing (Lester et al., 2014).
The symbol Et represents the conditional expectations operator, associated to the reference
probability distribution for εt+1. The likelihood ratio mt+1 between a distorted density ver-
sus the reference density allows to perform a change in the probability measure. The symbol
Et denotes the distorted conditional expectations operator, then Et[mt+1V (dt+1, kt+1, at+1)] =
Et[V (dt+1, kt+1, at+1)].
The penalty parameter θ measures the decision makers’ concern about robustness to misspec-
ification. This parameter enters the value function multiplying the relative entropy of distortion.
Hence if an alternative probability model is particularly unfavorable in terms of future expected
utility, it may not solve the minimization due to the countervailing effect of relative entropy. The
penalty has a lower bound θ, called neurotic breakdown point by Hansen and Sargent (2008). Below
this point it is useless to seek more robustness; the minimizing agent is sufficiently unconstrained
that he can push the criterion function to minus infinity. On the other hand, if θ goes to infinity,
the concern for model misspecification vanishes, so that the agent’s preferences are characterized
only by classical risk aversion. In the inner minimization the objective function is convex in mt+1
and the constraint is linear which allows to find the solution (as discussed by Hansen and Sargent
2008). The need for tractability of the minimization leads to the choice of relative entropy to
measure model discrepancies.
The solution of the minimization characterizes the worst-case distortion11:
m∗t+1 =exp
(−V (dt+1,kt+1,at+1)
θ
)
Et
[exp
(−V (dt+1,kt+1,at+1)
θ
)] (2.8)
8Schmitt-Grohe and Uribe (2003) and Schmitt-Grohe and Uribe (2017) discuss several ways to close SOE modelsand enable their convergence to a meaningful steady-state equilibrium. This equation is one of the different alternativesof model closure and is very popular in international macroeconomics.
9Following Lester et al. (2014) and Basu and Bundick (2017).10If x ∼ N(µ, σ2), then if X = ex the expected value of X is a function of the variance of x, E(X) = eµ+
σ2
2 .11The worst probabilistic scenario puts larger probability weights on bad productivity shocks.
CHAPTER 2. WELFARE COST OF MODEL UNCERTAINTY IN A SOE 11
Substituting m∗ into the original problem implies the following Bellman equation for the household
problem:
V (dt, kt, at) = maxct,ht,kt+1,dt+1
u(ct, ht)− βθ logEt exp
(−V (dt+1, kt+1, at+1)
θ
)(2.9)
s.t. dt+1 = (1 + rt)dt − atkαt h1−αt + ct + kt+1 − (1− δ)kt +
φ
2(kt+1 − kt)2
at+1 = (1− ρ)ass + ρat + ηεt+1
where the first restriction is the households’ budget constraint after replacing equations (2.3) to
(2.4) in equation (2.2).
The first-order conditions associated to the household’s maximization problem are:
λt = β(1 + rt+1)Et{m∗t+1λt+1
}= β(1 + rt+1)Et {λt+1} (2.10)
λt(1 + φ(kt+1 − kt)) = βEt{m∗t+1λt+1
[αat+1k
α−1t+1 h
1−αt+1 + 1− δ + φ(kt+2 − kt+1)
]}
= βEt{λt+1
[αat+1k
α−1t+1 h
1−αt+1 + 1− δ + φ(kt+2 − kt+1)
]}(2.11)
− uh(ct, ht) = λt(1− α)atkαt h−αt (2.12)
where λt is the marginal utility of consumption (λt = uc(ct, ht)), (2.10) is the Euler equation, (2.11)
is the first-order condition related to capital accumulation, and (2.12) is the first-order condition for
labor supply. In equations (2.10) and (2.11), as opposed to expected utility models, the equilibrium
conditions under multiplier preferences include the value function itself. Households are assumed
to be identical; therefore the average debt in equilibrium is equal to the households’ level of debt,
dt = dt.
2.3 Model solution
I start by referring to the perturbation approach to solve the model presented in the previous
section. Perturbation algorithms build a Taylor series expansion of the agents’ decision rules. I
implement a second-order approximation because the standard linearization method is useless for
this model, as discussed above in section 1. Perturbation methods were introduced by Judd and
Guu (1992) and intuitively explained by Schmitt-Grohe and Uribe (2004). Perturbation methods
are very fast and, despite their local nature, highly accurate for a large range of values of the state
variables (Arouba et al., 2006). Caldara et al. (2012) present an algorithm to use perturbation
methods, extending the work of Schmitt-Grohe and Uribe (2004) for recursive utility, in particular,
EZW preferences. Finally, Bidder and Smith (2012b) introduce Caldara et al. (2012) algorithm in
to the robust control literature. I adopt this method to solve my model.
CHAPTER 2. WELFARE COST OF MODEL UNCERTAINTY IN A SOE 12
2.3.1 Second-order approximation of the value function
The first step to calculate the second-order approximation of the value function is to write the
process of aggregate productivity in terms of a perturbation parameter χ, in the following way:
at+1 = (1− ρ)ass + ρat + χηεt+1 (2.13)
When the perturbation parameter value is set at χ = 1 (χ = 0), the model is stochastic (deter-
ministic). The parameter scaling the variance of the shocks is included in the set of vector of state
variables: st = (dt, kt, at;χ).
The second step is specifying the model’s equilibrium conditions augmented by the definition
of the value function, considering that all endogenous variables are functions of state variables:
V (st) = u(ct(st), ht(st))− βθ logEt exp
[−V ((st+1))
θ
](2.14)
λt(st) = β(1 + r∗t+1 + ψ(edt+1(st)−d − 1))Et
{exp(−Vt+1(st+1)
θ )
Et exp(−Vt+1(st+1)θ )
λt+1(st+1)
}(2.15)
λt(st)(1 + φ(kt+1(st)− kt)) = βEt
{exp(−Vt+1(st+1)
θ )
Et exp(−Vt+1(st+1)θ )
λt+1(st+1)
·[αat+1(st)kt+1(st)
α−1ht+1(st+1)1−α + 1− δ + φ(kt+2(st+1)− kt+1(st))]}
(2.16)
− uh(ct(st), ht(st)) = λt(st)(1− α)atkαt ht(st)
−α (2.17)
λt(st) = uc(ct(st), ht(st)) (2.18)
dt+1(st) = (1 + r∗t + ψ(edt−d − 1))dt − atkαt ht(st)1−α + ct(st) + kt+1(st)− (1− δ)kt
+φ
2(kt+1(st)− kt)2 (2.19)
at+1(st) = (1− ρ)ass + ρat + χηεt+1 (2.20)
The third step is to take the first derivatives of (2.14) to (2.19) with respect to the states st =
(dt, kt, at;χ) and evaluate them at the non-stochastic steady state. This leads to 24 equations
(6 equilibrium conditions times 4 state variables) and 24 unknowns (the first derivatives of the
6 endogenous ct, dt+1, ht, kt+1, λt and Vt variables with respect to the states evaluated at the
non-stochastic steady state). After solving the system of equations, the next step is to take the
derivatives of the first derivatives of (2.14) to (2.19) again with respect to the states: This step
arrives at a new system of 96 equations (6 first derivatives of the equilibrium conditions times
4 state variables) and 96 unknowns (the second derivatives of the 6 endogenous variables). The
results are reported in Appendix A.2.
Since the equilibrium conditions depend on the value function, it is necessary to derive its
approximation around the non-stochastic steady state, which allows to compute welfare calculations.
CHAPTER 2. WELFARE COST OF MODEL UNCERTAINTY IN A SOE 13
In order to simplify the exposition, I only report the approximation of the value function; but the
rest of the policy functions could be determined following the same procedure.
The second-order approximation of the value function is:
V (dt, kt, at;χ) 'Vss + Vd,ss(dt − dss) + Vk,ss(kt − kss) + Va,ss(at − ass) + Vχ,ssχ (2.21)
+ Vdk,ss(dt − dss)(kt − kss) + Vda,ss(dt − dss)(at − ass) + Vdχ,ss(dt − dss)χ
+1
2Vkk,ss(kt − kss)2 + Vka,ss(kt − kss)(at − ass) + Vkχ,ss(kt − kss)χ
+1
2Vaa,ss(at − ass)2 + Vaχ,ss(at − ass)χ+
1
2Vχχ,ssχ
2
Each term Vi,ss and Vij,ss is a scalar equal to the corresponding first and second-order derivative
of the value function for i, j = {d, k, a;χ}, evaluated at the non-stochastic steady state. The value
function evaluated at the non-stochastic steady state, V (dss, kss, ass; 0) = Vss. Hence, evaluating
the approximation of the value function at the non-stochastic steady state assuming χ = 1, yields
the following:
V (dss, kss, ass, 1) = Vss +1
2Vχχ,ss (2.22)
Van Binsbergen et al. (2009), show that Vχχ,ss is the only coefficient that is affected by uncertainty
aversion at a second-order approximation and that Vχχ,ss 6= 0. Vχχ,ss reflects the change in the
value function when the variance of productivity is η instead of zero.
For the robust SOE-RBC model described by equations (2.14) to (2.20), I derive the following
equation for the coefficient Vχχ,ss, the second derivative of the value function with respect to the
coefficient that scales the variance of productivity shocks (this coefficient is the only that depends
on robustness parameter θ):
Vχχ,ss =βη2
1− β
[Vaa,ss −
V 2a,ss
θ
](2.23)
2.3.2 Welfare analysis
The value function perturbation reflects that, up to a first-order approximation, the policy
functions of the model with multiplier preferences are equivalent to the expected utility model
considering the same instantaneous utility. But, at a second-order approximation, the value function
approximation differs between the two models by a term that is a function of the parameter that
governs robustness and the variance of the productivity shock. If the parameter that governs
robustness (θ) goes to infinity, reflecting that the concern for model misspecification goes to zero,
the second part of equation (2.23) goes to zero, which reflects the case when the household is only
risk averse. Therefore, expression (2.23) is the sum of two components: V riskχχ,ss is related to risk and
CHAPTER 2. WELFARE COST OF MODEL UNCERTAINTY IN A SOE 14
V robustχχ,ss is related to model uncertainty.
V riskχχ,ss =
βη2
1− βVaa,ss (2.24)
V robustχχ,ss = − βη2
1− βV 2a,ss
θ(2.25)
The term βη2
1−β that appears in both equations (2.24) and (2.25) is unambiguously positive. Hence
the sign of V riskχχ,ss depends only on Vaa,ss. As discussed by van Binsbergen et. al (2009), the latter
term has an ambiguous sign in a real business cycle model with endogenous labor and capital.
Cho et al. (2015) have shown that considering production economies under classical uncertainty
generates two effects on welfare: the fluctuation effect and the mean effect. The fluctuation effect
is always negative for the welfare of risk averse agents. The mean effect may increase or reduce
welfare, depending on the calibration of the model’s parameters. Therefore the total effect of risk
on the value function is ambiguous a priori.
On the other hand, the sign of the effect of the concern for model uncertainty V robustχχ,ss is unam-
biguously negative and its absolute value is decreasing in θ. A larger θ implies a smaller concern for
model misspecification. Accordingly, the overall effect of risk and model uncertainty on the value
function is ambiguous12.
It is important to note that the analytical decomposition and combination of the effects of the
two types of uncertainty considered here - classical risk and model uncertainty - and the resulting
ambiguous effect on overall welfare have not been derived in the previous literature.
2.3.3 Welfare cost calculations
This section presents calculations of the welfare cost of classical and model uncertainty, reflected
in consumption equivalent units. This implies computing the percentage loss in long-term mean
consumption (compensating variation) τ that would make the household indifferent between con-
suming (1−τ)css per period under certainty and css under uncertainty, where css is the steady-state
value of consumption. As shown by van Binsbergen et al. (2009), the coefficient Vχχ,ss of the value
function approximation could be used to measure the welfare cost of uncertainty. The term 12Vχχ,ss
represents how much indirect utility changes when the variance of the productivity shock is η2
instead zero.
For the robust SOE-RBC model developed in this paper, equation (2.23) decomposes the effects
of volatility and doubts on total welfare effect, deriving a closed-form solution for the cost to
face well-understood stochastic consumption fluctuations and the cost to face uncertainty about
the stochastic representation of productivity shocks. For welfare comparison, this expression is
transformed into consumption equivalent units. In order to do this, I introduce an explicit form of
households’ instantaneous utility function, which is due to by Greenwood, Hercowitz, and Huffman
12The intertemporal discount factor β and the standard deviation of productivity shocks η amplify the overall effectof risk and robustness on welfare in absolute value.
CHAPTER 2. WELFARE COST OF MODEL UNCERTAINTY IN A SOE 15
(1988):
u(ct, ht) =
(ct − hωt
ω
)1−σ− 1
1− σ (2.26)
Greenwood, Hercowitz, and Huffman (GHH) preferences, with a wage elasticity of labor supply
equal to 11−ω , have the advantage that the labor supply is insensitive to wealth effects because it
is independent of the level of consumption. This prevents persistent shocks affecting the level of
employment13. The parameter σ measures risk aversion, and its reciprocal is the intertemporal
elasticity of substitution.
Considering GHH preferences (2.26), I compute τ the welfare loss due to aggregate uncertainty
from the following:
(css − hωss
ω
)1−σ− 1
1− σ +1
2Vχχ,ss =
(css(1− τ)− hωss
ω
)1−σ− 1
1− σ (2.27)
Replacing Vχχ,ss from (2.23) into (2.27) an solving for τ :
τ =
(css − hωt
ω
)
css− 1
css
(βη2
2(1− σ)
[Vaa,ss −
V 2a,ss
θ
]+
(css −
hωssω
)1−σ) 11−σ
(2.28)
It is not possible to continue advancing analytically as the coefficient Vaa,ss can only be determined
numerically, nevertheless, Va,ss as all first order coefficients of value function’s approximation are
determined analytically and presented in Appendix A.2.
2.4 Model calibration and numerical calculations
This section presents numerical calculations for the welfare cost by means of consumption
equivalent units in order to perform welfare comparisons of the model’s two sources of uncertainty:
risk and model uncertainty. The calibration of the main parameters of the SOE-RBC model follows
Schmitt-Grohe and Uribe (2003) and is summarized in Table 2.114:
Table 2.1: Calibration of model parameters
δ r∗ α d ω φ ψ ρ η β ass
0.1 0.04 0.32 0.7442 1.455 0.028 0.000742 0.42 0.0129 0.9615 1
where δ is the depreciation rate, r∗ is the international interest rate, α is the capital share, d is
steady-state debt, ω is the wage elasticity of labor supply, φ is the capital adjustment cost coefficient,
ψ is the sensitivity of sovereign risk to debt, ρ reflects the persistence of productivity shocks, η is
the volatility of productivity shocks, β is the subjective discount factor, and ass is the steady-state
value of productivity.
13By eliminating the wealth effect, these preferences raise the likelihood that welfare improves with uncertainty, asshown by Lester et al. (2014).
14Parameter calibration by Schmitt-Grohe and Uribe (2003) follows Mendoza (1991) original calibration for Canada.
CHAPTER 2. WELFARE COST OF MODEL UNCERTAINTY IN A SOE 16
For the key robustness parameter θ, for my extended model, I use three different values, which
span the range of possible values. The first value is infinity, which represents the classical uncertainty
case, where the concern for model misspecification is absent. The second is θ = 8, which comes
from Bidder and Smith (2012b), who estimated this value by using error-detection probabilities15
that express how difficult it is for a decision-maker, with limited data, to distinguish between the
worst-case probabilistic scenario and the benchmark model. The third value is set at θ = 1.2, to
consider a case of extremely high concern for model misspecification. This is the smallest value for
which model convergence is still achieved.
For the risk-aversion coefficient σ, I also use three different values: σ = 1 corresponding to
logarithmic preferences; σ = 2, used by Schmitt-Grohe and Uribe (2003); and σ = 5, which is a
high value for risk aversion. Appendix A.3 presents welfare calculations for a broader set of values
for the risk aversion parameter to check for consistency of the results.
The following Table 2.2 reports the percentage loss in long-term mean consumption, that makes
the household indifferent between consuming (1 − τ)css per period under certainty and css under
uncertainty, for different values of risk aversion (σ). If the value of τ , determined by equation (2.28),
is positive (negative), facing uncertainty implies a welfare cost (premium) for the representative
agent. The table reports the compensating variation τ for nine combinations of values for the
agents’ degree of risk aversion σ and their concern for robustness θ.
Table 2.2: Welfare loss estimations due to aggregate uncertainty, for different parameter values ofrisk aversion and concern for model uncertainty (Percentage loss in long-term mean consumption,τ)
Model Risk Aversion
σ =1 σ =2 σ =5
Risk (θ =∞) -0.0083 -0.0042 0.008
Risk + Model Uncertainty (θ = 8) 0.0048 0.0282 0.4082
Risk + Model Uncertainty (θ = 1.2) 0.0786 0.2021 2.3322
Let’s consider first the case when households are only risk averse (first line in Table 2.2). At
relatively low levels of risk aversion (σ = 1 and 2), the agent reaps a welfare premium. But when
risk aversion is larger (in the table, at a value of σ = 5), the presence of risk is welfare deteriorating.
This reflects the well-established result that risk has an ambiguous effect on welfare.
Now consider adding the households’ concern about model misspecification, which represents
an unambiguous welfare loss (second and third lines in the table). The presence of this welfare loss
turns the welfare gain, obtained at relatively low levels of risk aversion (σ = 1 and 2), into an overall
welfare loss. And at a high level of risk aversion (σ= 5), the overall welfare loss is significantly
raised by the concern about model specification.
The latter is observed even at a moderate degree of concern for model uncertainty (θ = 8, line
2). At a very high level of concern for model uncertainty (θ = 1.2, line 3), the overall welfare
15This methodology is also applied by Hansen and Sargent (2008).
CHAPTER 2. WELFARE COST OF MODEL UNCERTAINTY IN A SOE 17
costs of uncertainty are one order of magnitude larger than at a moderate level. When a high
level of concern for model uncertainty (θ= 1.2) is combined with high risk aversion (σ= 5), the
overall welfare loss due to uncertainty reaches a very high level, equivalent to 2.3% of long-term
consumption.
In Figure 1, I expand the calculations for the compensating variation of consumption reported
in Table 2 for a wider range of risk aversion values, extending from σ = 0 to σ= 5. The three
lines are drawn for the corresponding three values of θ that I set above. The figure reflects how the
overall welfare loss due to uncertainty grows exponentially with risk aversion and, in particular,
with the concern about model uncertainty.
Figure 2.1: Welfare loss estimations due to aggregate uncertainty, for different parameter valuesof risk aversion and concern for model uncertainty; no financial frictions case (Percentage loss inlong-term mean consumption, τ)
Up to here I have considered a low sensitivity of the debt risk premium to the level of sovereign
debt, reflected in ψ=0.000742, following Schmitt-Grohe and Uribe (2003). Now I consider a high
value for the debt sensitivity parameter, consistent with the existence of financial frictions. Eric-
son and Liu (2012) examine the welfare cost of business cycles introducing financial frictions in a
SOE with a risk-averse agent, without any other concern for uncertainty. They find that produc-
CHAPTER 2. WELFARE COST OF MODEL UNCERTAINTY IN A SOE 18
tivity shocks are welfare improving in the absence of financial frictions but they become welfare
deteriorating under financial frictions.
For the financial frictions case, I set ψ at a value of 5, which is the median value estimated by
Garcıa-Cicco et al. (2010). Figure 2 presents the welfare cost estimations for the financial frictions
case, analogous to Figure 1 where financial frictions were absent.
Figure 2.2: Welfare loss estimations due to aggregate uncertainty, for different parameter values ofrisk aversion and concern for model uncertainty; financial frictions case (Percentage loss in long-termmean consumption, τ)
Comparison of Figures 1 and 2 shows that in the absence of concern for model uncertainty (θ =∞),
the introduction of financial frictions reduces the positive effect of uncertainty at low levels of risk
aversion and raises the negative effect diminished at high levels of risk aversion. This replicates
the results reported by Ericson and Liu (2012), in the context of their SOE-RBC model, limited
to classical risk. My model adds model uncertainty to classical risk about productivity. Analyzing
my model without financial frictions, uncertainty becomes costly and the cost is increasing in the
level of concern for model misspecification. However, when adding financial frictions the effect of
CHAPTER 2. WELFARE COST OF MODEL UNCERTAINTY IN A SOE 19
uncertainty on the highest welfare cost by about 20%. The explanation for this result is due to the
link between consumption, foreign debt, and the sovereign risk premium. Model uncertainty gen-
erates an increase in precautionary savings, then sovereign risk declines, and the domestic interest
rate falls. The opportunity cost of consumption decreases, which softens the fall in consumption in
response to the increase in precautionary savings.
2.5 Conclusions
This paper contributes to the literature on the welfare cost of business-cycle fluctuations in a
SOE by introducing model uncertainty. Households face two types of uncertainty: classical uncer-
tainty about the state of future aggregate productivity and model uncertainty about the probability
model that describes the movements of productivity. The characterization of the small open econ-
omy follows the canonical SOE-RBC of Schmitt-Grohe and Uribe (2003), augmented by multiplier
preferences (Hansen and Sargent 2008). These preferences correspond to a non-expected utility
approach that reflects a concern for model misspecification (model uncertainty). Households are
endowed with a reference probability model about the stochastic variable (aggregate productivity),
but they distrust its accuracy, considering the possibility that the model is misspecified in a way
that is difficult to detect statistically. Barillas et al. (2009) and Ellison and Sargent (2015) highlight
that if agents care about robustness to model misspecification, business cycle fluctuations are more
costly in terms of agents’ welfare than if only risk aversion is considered.
Using perturbation methods, the paper derives a closed-form solution for the combined welfare
effect of the two sources of uncertainty. While consumption fluctuations have an ambiguous effect
on welfare (as shown by Cho et al., 2015 and Lester et al., 2014), the additional model uncer-
tainty is unambiguously welfare-deteriorating. Hence the overall effect of uncertainty on welfare is
ambiguous, depending on consumer preferences and model parameters. Then, the paper provides
numerical results for the welfare effects of uncertainty. The simulations show that when classical
risk has a positive effect on welfare (due to a low risk aversion), addition of model uncertainty can
reverse the sign of the overall welfare effect. The total effect of uncertainty, including both sources,
is shown to be negative for a wide range of parameter values.
The relevance of this paper’s results lies in the broad discussion about macroeconomic stabiliza-
tion, as highlighted by Lucas (1987, 2003). If the welfare cost of business cycles is negligible or if
households prefer economic uncertainty, there is no space for counter-cyclical policy in small open
economies. My first result for numerical simulations shows that for slightly risk-averse households,
productivity shocks are in general welfare improving. However, if they are intensely risk averse,
the welfare cost of productivity risk is estimated at 0.007% of long-term consumption, similar to
the value of 0.008% found by Lucas (1987). The second finding is that by introducing model un-
certainty, even when the agent is only mildly risk averse, the overall welfare cost of uncertainty
becomes welfare-deteriorating. The third result, based on numerical calculations, shows that the
total effect of uncertainty on welfare depends on the interaction of risk and model uncertainty.
Welfare costs are increasing in both risk aversion and the concern for model uncertainty. Consider-
CHAPTER 2. WELFARE COST OF MODEL UNCERTAINTY IN A SOE 20
ing a moderately high degree of risk aversion and a reasonable concern for model uncertainty, the
overall effect of uncertainty on welfare amounts to a loss of 0.4% of long-term consumption, which
is two orders of magnitude higher than the finding by Lucas. If I consider the case of an agent
with very high concern for model misspecification, the welfare cost reaches a staggering 2.3% of
long-term consumption. Finally, I consider the existence of financial frictions, reflected in a high
debt sensitivity of the sovereign risk premium, which has been proven to increase the welfare cost of
productivity shocks in a SOE, for the case of classical risk aversion (Ericson and Liu 2012). In a risk
aversion context the largest simulated welfare cost of business cycles is 0.013, which is significantly
smaller than the largest value obtained when including preferences for robustness. Therefore, with
or without financial frictions households with a preference for robustness are willing to sacrifice a
lot to live in an economy without fear of misspecification.
Bibliography
[1] Alvarez, F. and U. Jermann (2004): Using Asset Prices to Measure the Cost of Business Cycles.
Journal of Political Economy, vol. 112(6), 1223-1256.
[2] Arouba, S. B., J. Fernandez-Villaverde and J. F. Rubio-Ramırez (2006): Comparing Solution
Methods for Dynamic Equilibrium Economies. Journal of Economic Dynamics and Control, vol.
30(12), 2477-2508.
[3] Arrow, K. (1951): Alternative Approaches to the Theory of Choice in Risk-Taking. Economet-
rica, vol. 19, 404-437.
[4] Barillas, F., L. P. Hansen, and T. Sargent (2009): Doubts or variability? Journal of Economic
Theory, vol. 144(6), 2388-2418.
[5] Basu, S. and B. Bundick (2017): Uncertainty Shocks in a Model of Effective Demand. Econo-
metrica, vol. 85(3), 937-958.
[6] Bidder, R. and Smith M. (2012a): Robust Control in a Nonlinear DSGE Model. Manuscript,
New York University.
[7] Bidder, R. and M. Smith (2012b): Robust Animal Spirits. Journal of Monetary Economics, vol.
59(8), 738-750.
[8] Bidder, R. and M. Smith (2015): Doubts and Variability: A Robust Perspective on Exotic
Consumption Series. Working Paper Series from Federal Reserve Bank of San Francisco, No
2013-28.
[9] Van Binsbergen, J.H., J. Fernandez-Villaverde, R.S.J. Koijen, and J. Rubio-Ramırez (2009):
Likelihood estimation of DSGE models with Epstein-Zin preferences. Manuscript, University of
Pennsylvania.
[10] Caldara, D., J. Fernandez-Villaverde, J. Rubio-Ramı rez, and W. Yao (2012): Computing
DSGE Models with Recursive Preferences and Stochastic Volatility. Review of Economic Dy-
namics, vol. 15(2), 188-206.
[11] Cho, J. and T. Cooley (2000): Business Cycle Uncertainty and Economic Welfare. Manuscript,
New York University.
21
BIBLIOGRAPHY 22
[12] Cho, J., T. Cooley, and H. Kim (2015): Business Cycle Uncertainty and Economic Welfare.
Review of Economics Dynamics, vol. 18(2), 185-200.
[13] Dolmas, J. (1998): Risk Preferences and the Welfare Cost of Business Cycles. Review of Eco-
nomic Dynamics, vol. 1(3), 646-676.
[14] Ellison, M. and T. J. Sargent (2015): Welfare Cost of Business Cycles with Idiosyncratic Con-
sumption Risk and Preferences for Robustness. American Economic Journal: Macroeconomics,
vol. 7(2), 40-57.
[15] Epaulard, A. and A. Pommeret (2003): Recursive Utility, Endogenous Growth, and the Welfare
Cost of Volatility. Review of Economic Dynamics, vol. 6, 672-684.
[16] Epstein, L. G. and S. E. Zin (1989): Substitution, Risk Aversion, and the Temporal Behavior
of Consumption and Asset Returns: a Theoretical Framework. Econometrica, vol. 57, 937-969.
[17] Ericson, R. and X. Liu, (2012): Welfare Effect of Productivity Shocks and Policy Implications
in a Small Open Economy. Perspective on Globe Development and Technology, vol. 11, 290-319.
[18] Garcıa-Cicco, J., R. Pancrazi and M. Uribe (2010): Real business cycles in emerging countries?
American Economic Review, vol. 100(5), 2510-31.
[19] Greenwood, J., Z. Hercowitz, and G. Huffman (1988): Investment, Capacity Utilization, and
the Real Business Cycle. American Economic Review, vol. 78(3), 402-417.
[20] Hansen, L. P. and M. Marinacci (2016): Ambiguity Aversion and Model Misspecification: An
Economic Perspective. Statististical Science, vol.31(4), 511–515.
[21] Hansen, L. P. and T. J. Sargent (2001): Robust Control and Model Uncertainty. American
Economic Review, vol. 91, 60-66.
[22] Hansen, L. P. and T. J. Sargent (2008): Robustness. Princeton University Press.
[23] Hansen, L. P. and T. J. Sargent (2010): Wanting Robustness in Macroeconomics. Handbook
of Monetary Economics, vol. 3, chapter 20, 1097-1157.
[24] Judd, K. L. and S.M. Guu (1992): Perturbation Solution Methods for Economic Growth
Models. Economic and Financial Modelling with Mathematica, Springer-Verlag.
[25] Knight, F. H. (1921): Risk, Uncertainty and Profit. Houghton Mifflin.
[26] Krusell, P. and A. Smith (1999): On the Welfare Effects of Eliminating Business Cycles. Review
of Economic Dynamics, vol. 2(2), 245-272.
[27] Lester, R., M. Pries, and E. Sims (2014): Volatility and Welfare. Journal of Economic Dynamic
& Control, vol. 38, 17-36.
BIBLIOGRAPHY 23
[28] Levin, A., D. Lopez-Salido, N. Edward, and T. Yun (2008): Macroeconometric Equivalence,
Microeconomic Dissonance, and the Design of Monetary Policy. Journal of Monetary Economics,
vol. 55(Supplemen), 48-62.
[29] Lucas, Jr., R. E. (1987): Models of Business Cycles. Oxford: Basil Blackwell.
[30] Lucas, Jr., R. E. (2003): Macreconomic Priorities. American Economic Review: Papers and
Proceeding, vol 93, 1-14.
[31] Maccheroni, F., M. Marinacci, and A. Rustichini (2006a): Ambiguity Aversion, Robustness,
and the Variational Representation of Preferences. Econometrica, vol. 74(6), 1447-1498.
[32] Maccheroni, F., M. Marinacci, A. and Rustichini (2006b): Dynamical Variational Preferences.
Journal of Economic Theory, vol. 128, 4-44.
[33] Obstfeld, M. (1994): Evaluating risky consumption paths: The Role of Intertemporal Substi-
tutability. European Economic Review, vol. 38(7), 1471-1486.
[34] Schmitt-Grohe, S. and M. Uribe (2003): Closing Small Open Economy Models. Journal of
International Economics, vol. 61, 163-185.
[35] Schmitt-Grohe, S. and M. Uribe (2004): Solving Dynamic General Equilibrium Models Using
a Second-order Approximation to the Policy Function. Journal of Economic Dynamics and
Control, vol. 28(4), 755-775.
[36] Tallarini, T. D. (2000). Risk-Sensitive Real Business Cycles. Journal of Monetary Economics,
vol. 45, 507-532.
[37] Weil, P. (1990): Non-Expected Utility in Macroeconomics. Quarterly Journal of Economics,
vol. 1, 29- 42.
[38] Xu, S. (2017): Volatility Risk and Economic Welfare. Journal of Economic Dynamics and
Control, vol. 80, 17-33.
Chapter 3
Identifying Quantum Structures in
the Ellsberg Paradox
Abstract
Empirical evidence has confirmed that quantum effects occur frequently also outside the micro-
scopic domain, while quantum structures satisfactorily model various situations in several areas of
science, including biological, cognitive and social processes. In this paper, we elaborate a quantum
mechanical model which faithfully describes the Ellsberg paradox in economics, showing that the
mathematical formalism of quantum mechanics is capable to represent the ambiguity present in
this kind of situations, because of the presence of contextuality. Then, we analyze the data col-
lected in a concrete experiment we performed on the Ellsberg paradox and work out a complete
representation of them in complex Hilbert space. We prove that the presence of quantum structure
is genuine, that is, interference and superposition in a complex Hilbert space are really necessary
to describe the conceptual situation presented by Ellsberg. Moreover, our approach sheds light
on ‘ambiguity laden’ decision processes in economics and decision theory, and allows to deal with
different Ellsberg-type generalizations, e.g., the Machina paradox.
3.1 Introduction
Traditional approaches in economics follow the hypothesis that agents’ behavior during a deci-
sion process is mainly ruled by expected utility theory (EUT) [1, 2]. Roughly speaking, in presence
of uncertainty, decision makers choose in such a way that they maximize their expected utility.
Notwithstanding its mathematical tractability and predictive success, the structural validity of
EUT at the individual level is questionable. Indeed, systematic empirical deviations from the
predictions of EUT have been observed which are usually referred to as paradoxes [3, 4].
EUT was formally elaborated by von Neumann and Morgenstern [1]. They presented a set
of axioms that allow to represent decision–maker preferences over the set of acts (functions from
the set of states of the nature into the set of consequences) by a suitable functional Epu(.), for
24
CHAPTER 3. IDENTIFYING QUANTUM STRUCTURES IN THE ELLSBERG PARADOX25
Act red yellow black
f1 12$ 0$ 0$
f2 0$ 0$ 12$
f3 12$ 12$ 0$
f4 0$ 12$ 12$
Table 3.1: The payoff matrix for the Ellsberg paradox situation.
some Bernoulli utility function u on the set of consequences and an objective probability measure
p on the set of states of the nature. An important aspect of EUT concerns the treatment of
uncertainty. Knight had pointed out the difference between risk and uncertainty reserving the
term risk for situations that can be described by known (or physical) probabilities, and the term
uncertainty to refer to situations in which agents do not know the probabilities associated with
each of the possible outcomes[5]. Von Neumann and Morgenstern modeling did not contemplate
the latter possibility, since all probabilities are objectively, i.e. physically, given in their scheme. For
this reason, Savage extended EUT allowing agents to construct their own subjective probabilities
when physical probabilities are not available [2]. According to Savage’s model, the distinction
proposed by Knight seems however irrelevant. Ellsberg instead showed that Knightian’s distinction
is empirically meaningful [3]. In particular, he presented the following experiment. Consider one
urn with 30 red balls and 60 balls that are either yellow or black, the latter in unknown proportion.
One ball will be drawn from the urn. Then, free of charge, a person is asked to bet on one of the
acts f1, f2, f3 and f4 defined in Table ??. When asked to rank these gambles most of the persons
choose to bet on f1 over f2 and f4 over f3. This preference cannot be explained by EUT. Indeed,
individuals’ ranking of the sub–acts [12 on red; 0 on black] versus [0 on red; 12 on black] depends
upon whether the event yellow yields a payoff of 0 or 12, contrary to what is suggested by the
Sure–Thing principle, an important axiom of Savage’s model.1 Nevertheless, these choices have a
direct intuition: f1 offers the 12 prize with an objective probability of 1/3, and f2 offers the same
prize but in an element of the subjective partition {black, yellow}. In the same way, f4 offers the
prize with an objective probability of 2/3, whereas f3 offers the same payoff on the union of the
unambiguous event red and the ambiguous event yellow. Thus, in both cases the unambiguous bet
is preferred to its ambiguous counterpart, a phenomenon called ambiguity aversion by Ellsberg.
Many extensions of EUT have been worked out to cope with Ellsberg–type preferences, which
mainly consist in replacing the Sure–Thing Principle by weaker axioms. We briefly summarize the
most known, as follows.
(i) Choquet expected utility. This model considers a subjective non–additive probability (or,
capacity) over the states of nature rather than a subjective probability. Thus, decision–makers could
underestimate or overestimate probabilities in the Ellsberg experiment, and ambiguity aversion is
1The Sure–Thing principle was stated by Savage by introducing the businessman example, but it can be presentedin an equivalent form, the independence axiom, as follows: if persons are indifferent in choosing between simplelotteries L1 and L2, they will also be indifferent in choosing between L1 mixed with an arbitrary simple lottery L3
with probability p and L2 mixed with L3 with the same probability p.
CHAPTER 3. IDENTIFYING QUANTUM STRUCTURES IN THE ELLSBERG PARADOX26
equivalent to the convexity of the capacity (pessimistic beliefs) [6].
(ii) Max–Min expected utility. The lack of knowledge about the states of nature of the decision–
maker cannot be represented by a unique probability measure but, rather, by a set of probability
measures. Then, an act f is preferred to g iff minp∈P Epu(f) > minp∈P Epu(f), where P is a
convex and closed set of additive probability measures. Ambiguity aversion is represented by the
pessimistic beliefs of the agent which takes decisions considering the worst probabilistic scenario
[7].
(iii) Variational preferences. In this dynamic generalization of the Max–Min expected utility,
agents rank acts according to the criterion infp∈4{Epu(f)+c(p)}, where c(p) is a closed and convex
penalty function associated with the probability election [8].
(iv) Second order probabilities. This is a model of preferences over acts where the decision–
maker prefers act f to act g iff Eµφ(Epu(f)) > Eµφ (Epu(g)), where E is the expectation operator,
u is a von Neumann–Morgenstern utility function, φ is an increasing transformation, and µ is a
subjective probability over the set of probability measures p that the decision–maker thinks are
feasible. Ambiguity aversion is here represented by the concavity of the transformation φ [9].
Despite approaches (i)–(iv) have been widely used in economic and financial modeling, none of
them is immune of critics [4, 10]. And, worse, none of these models can satisfactorily represent
more general Ellsberg–type situations (e.g., the Machina paradox [4, 11]). As a consequence, the
construction of a unified perspective representing ambiguity is still an unachieved goal in economics
and decision making.
We have recently inquired both conceptually and structurally into the above approaches gen-
eralizing EUT [6, 7, 8, 9] to cope with ambiguity. The latter is defined as a situation without a
probability model describing it as opposed to risk, where a classical probability model on a σ–
algebra of events is presupposed. The generalizations in (i)–(iv) consider more general structures
than a single classical probability model on a σ–algebra. We are convinced that this is the point:
ambiguity, due to its contextuality, structurally need a non–classical probability model. To this
end we have elaborated a general framework, based on the notion of contextual risk and inspired
by the probability structure of quantum mechanics, which is intrinsically different from a classical
probability on a σ-algebra, the set of events is indeed not a Boolean algebra [13, 14, 15, 16]. In-
spired by this approach, we work out in the present article a complete mathematical representation
of the Ellsberg paradox situation (states, payoffs, acts, preferences) in the standard mathemati-
cal formalism of quantum mechanics, hence by using a complex Hilbert space, and representing
the probability measures by means of projection valued measures on this complex Hilbert space
[17] (Sect. 3.2). This analysis leads us to claim that the structure of the probability models is
essentially different from the ones of known approaches – projection valued measures instead of
σ–algebra valued measures [12]. But, more important, also the way in which states are repre-
sented in quantum mechanics, i.e. by unit vectors of the Hilbert space, introduces a fundamentally
different aspect, coping both mathematically and intuitively with the notion of ambiguity as intro-
duced in economics. Successively, we analyze the experimental data we collected in a statistically
relevant experiment we performed, where real decision–makers were asked to bet on the different
CHAPTER 3. IDENTIFYING QUANTUM STRUCTURES IN THE ELLSBERG PARADOX27
acts defined by the Ellsberg paradox situation [13, 14] (Sect. 3.3). We show that our quantum
mechanical model faithfully represents the subjects’ preferences together with experimental statis-
tics. Furthermore, we describe the choices between acts f1/f2 and between acts f3/f4 by quantum
observables represented by compatible spectral families (Sect. 3.4). We finally prove that the re-
quirement of compatibility of the latter observables makes it necessary to introduce a Hilbert space
over complex numbers, namely that imaginary numbers are needed since our experimental data
cannot be modeled in a real Hilbert space, i.e. a vector space over real numbers only, in case we
want the observables representing our experiment to be compatible (Sect. 3.5). Complex numbers
in quantum theory stand for the quantum effect of interference, and indeed, we can identify in our
modeling how interference produces the typical Ellsberg deviation leading to measured data in our
experiment.
The results above strongly suggest that, more generally, ‘ambiguity laden’ situations can be
explained in terms of the appearance of typically quantum effects, namely, contextuality – our
quantum model is contextual, see the discussion on context in Sect. 3.2 –, superposition – we
explicitly use superposition to construct the quantum states representing the Ellsberg bet situations
in Sect. 3.2 – and interference – i.e. complex numbers –, and that quantum structures have the
capacity to mathematically deal with this type of situations. Hence, our findings naturally fit within
the growing ‘quantum interaction research’ which mainly applies quantum structures to cognitive
situations [18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28].
3.2 A quantum model in Hilbert space for the Ellsberg paradox
Here we present a Hilbert space representation of the Ellsberg situation and derive new results
elaborating the already obtained ones. But we first need to anticipate a discussion on the notions of
‘context’ and ‘contextual influence’ and how they are used in the present framework. The notion of
context typically denotes what does not pertain to the entity under study but that can interact with
it. In the foundations of quantum mechanics, context more specifically indicates the ‘measurement
context’, which influences the measured quantum entity in a stochastic way. As a consequence of
this contextual interaction, the state of the quantum entity changes, thus determining a transition
from potential to actual. A quantum mechanical context is represented by a self-adjoint operator
or, equivalently, by a spectral family. An analogous effect occurs in a decision process, where there
is generally a contextual influence (of a cognitive nature) having its origin in the way the mind of
the person involved in the decision, e.g., a choice between two bets, relates to the situation that is
the subject of the decision making, e.g., the Ellsberg situation. This is why, in our analysis of the
Ellsberg paradox, we use the definition and representation of context and contextual influence to
indicate the cognitive interaction taking place between the conceptual Ellsberg situation and the
human mind in a decision process. We are now ready to proceed with our quantum modeling.
To this end let us consider the situation illustrated in Tab. 1, Sect. 3.1. The simplest Hilbert
space that can supply a faithful modeling of the Ellsberg situation is the three dimensional complex
Hilbert space C3 whose canonical basis we denote by {|1, 0, 0〉, |0, 1, 0〉, |0, 0, 1〉}. We introduce the
CHAPTER 3. IDENTIFYING QUANTUM STRUCTURES IN THE ELLSBERG PARADOX28
model in different steps: we will see that in each of them, quantum structures enable a new
and satisfying way to model an aspect of ambiguity. At the last step it will be made clear that
modeling the agents’ decisions at a statistical level requires the full probabilistic apparatus of
quantum mechanics, that is, both states and measurements.
3.2.1 The conceptual Ellsberg entity
The first part of the model consists of the Ellsberg situation without considering neither the
different acts nor the person nor the bet to be taken. Hence it is the situation of the urn with 30
red balls and 60 black and yellow balls in unknown proportion (conceptual Ellsberg entity). Already
at this stage, the presence of ambiguity can be mathematically taken into account by means of
the quantum mechanical formalism. To this aim we introduce a quantum mechanical context e
and represent it by means of the family {Pr, Pyb}, where Pr is the one dimensional orthogonal
projection operator on the subspace generated by the unit vector |1, 0, 0〉, and Pyb is the two
dimensional orthogonal projection operator on the subspace generated by the unit vectors |0, 1, 0〉and |0, 0, 1〉. {Pr, Pyb} is a spectral family, since Pr ⊥ Pyb and Pr + Pyb = 1. Contexts, more
specifically measurement contexts, are represented by spectral families of orthogonal projection
operators (equivalently, by a self–adjoint operator determined by such a family) also in quantum
mechanics. Again in analogy with quantum mechanics, we represent the states of the conceptual
Ellsberg entity by means of unit vectors of C3. For example, the unit vector
|vry〉 = | 1√3eiθr ,
√2
3eiθy , 0〉 (3.1)
can be used to represent a state describing the Ellsberg situation. Indeed, the probability for ‘red’
in the state pvry represented by |vry〉 is
|〈1, 0, 0|vry〉|2 = 〈vry|1, 0, 0〉〈1, 0, 0|vry〉 =‖ Pr|vry〉 ‖2=1
3(3.2)
Moreover, the probability for ‘yellow or black’ in the state pvry represented by |vry〉 is
‖ Pyb|vry〉 ‖2= 〈0,√
2
3eiθy , 0|0,
√2
3eiθy , 0〉 =
2
3(3.3)
But this is not the only state describing the Ellsberg situation. For example, the unit vector
|vrb〉 = | 1√3eiφr , 0,
√2
3eiφb〉 (3.4)
also represents a state describing the Ellsberg situation. We thus denote the set of all states
describing the Ellsberg situation (Ellsberg state set) by
ΣElls = {pv : |v〉 = | 1√3eiθr , ρye
iθy , ρbeiθb〉 | 0 ≤ ρy, ρb, ρ2
y + ρ2b =
2
3} (3.5)
CHAPTER 3. IDENTIFYING QUANTUM STRUCTURES IN THE ELLSBERG PARADOX29
which is associated with a subset (not necessarily a linear subspace) of C3. If a state belongs to
ΣElls, this state delivers a quantum description of the Ellsberg situation, together with the context
e represented by the spectral family {Pr, Pyb} in C3.
3.2.2 Modeling acts and utility
In the second step of our construction, we describe the different acts f1, f2, f3 and f4. Here a
second measurement context g is introduced. The context g describes the ball taken out of the urn
and its color verified, red, yellow or black. Also g is represented by a spectral family of orthogonal
projection operators {Pr, Py, Pb}, where Pr is already defined, while Py is the orthogonal projection
operator on |0, 1, 0〉 and Pb is the orthogonal projection operator on |0, 0, 1〉. Thus, the probabilities
µr(g, pv), µy(g, pv) and µb(g, pv) of drawing a red ball, a yellow ball and a black ball, respectively,
in a state pv represented by the unit vector |v〉 = |ρreiθr , ρyeiθy , ρbeiθb〉 are
µr(g, pv) =‖ Pr|v〉 ‖2= 〈v|Pr|v〉 = ρ2r (3.6)
µy(g, pv) =‖ Py|v〉 ‖2= 〈v|Py|v〉 = ρ2y (3.7)
µb(g, pv) =‖ Pb|v〉 ‖2= 〈v|Pb|v〉 = ρ2b (3.8)
The acts f1, f2, f3 and f4 are observables in our modeling, hence they are represented by
self-adjoint operators, built on the spectral decomposition {Pr, Py, Pb}. More specifically, we have
F1 = 12$Pr (3.9)
F2 = 12$Pb (3.10)
F3 = 12$Pr + 12$Py (3.11)
F4 = 12$Py + 12$Pb = 12$Pyb (3.12)
Let us now analyze the expected payoffs and the utility connected with the different acts. For the
sake of simplicity, we identify here the utility with the expected payoff, which implies that we are
considering risk neutral agents. Consider an arbitrary state pv ∈ ΣElls and the acts f1 and f4. We
have
U(f1, g, pv) = 〈pv|F1|pv〉 = 12$ · 1
3= 4$ (3.13)
U(f4, g, pv) = 〈pv|F4|pv〉 = 12$ · 2
3= 8$ (3.14)
which shows that both these utilities are completely independent of the considered state of ΣElls, i.e.
they are ambiguity free. Consider now the acts f2 and f3, and again an arbitrary state pv ∈ ΣElls.
We have
U(f2, g, pv) = 〈pv|F2|pv〉 = 12$µb(g, pv) (3.15)
U(f3, g, pv) = 〈pv|F3|pv〉 = 12$(µr(g, pv) + µy(g, pv)) (3.16)
CHAPTER 3. IDENTIFYING QUANTUM STRUCTURES IN THE ELLSBERG PARADOX30
which shows that both utilities strongly depend on the state pv, due to the ambiguity the two acts
are confronted with. This can be significantly revealed by considering two extreme cases. Let pvryand pvrb be the states represented by the vectors |vry〉 and |vrb〉 in Eqs. (3.1) and (3.4), respectively.
These states give rise for the act f2 to utilities
U(f2, g, pvry) = 12$µb(g, pvry) = 12$ · 0 = 0$ (3.17)
U(f2, g, pvrb) = 12$µb(g, pvrb) = 12$ · 2
3= 8$. (3.18)
This shows that a state pvrb exists within the realm of ambiguity, where the utility of act f2 is
greater than the utility of act f1, and also a state pvry exists within the realm of ambiguity, where
the utility of act f2 is smaller than the utility of act f1. If we look at act f3, we find for the two
considered extreme states the following utilities
U(f3, g, pvry) = 12$(µr(g, pvry) + µy(g, pvry)) = 12$(1
3+
2
3) = 12$ (3.19)
U(f3, g, pvrb) = 12$(µr(g, pvrb) + µy(g, pvrb)) = 12$(1
3+ 0) = 4$. (3.20)
Analogously, namely the state pvry gives rise to a greater utility, while the state pvrb gives rise to a
smaller utility than the independent one obtained in act f4.
3.2.3 Decision making and superposition states
The third step of our modeling consists in taking directly into account the role played by
ambiguity. We proceed as follows [21].
We suppose that the two extreme states pvry and pvrb in Sect. 3.2.1 are relevant in the mind of
the person that is asked to bet. Hence, it is a superposition state of these two states that will guide
the decision process during the bet. Let us construct a general superposition state pvs of these two
states. Hence the vector |vs〉 representing pvs can be written as
|vs〉 = aeiα|vrb〉+ beiβ|vry〉 (3.21)
where a, b, α and β are chosen in such a way that 〈vs|vs〉 = 1, which means that
1 = a2 + b2 +2ab
3cos(β − α+ θr − φr) (3.22)
or, equivalently,
cos(β − α+ θr − φr) =3(1− a2 − b2)
2ab(3.23)
The amplitude of the state pvs with the first basis vector |1, 0, 0〉 is given by
〈1, 0, 0|vs〉 =a√3ei(α+φr) +
b√3ei(β+θr) (3.24)
CHAPTER 3. IDENTIFYING QUANTUM STRUCTURES IN THE ELLSBERG PARADOX31
Therefore, the transition probability is
|〈1, 0, 0|vs〉|2 =1
3(a2 + b2 + 3− 3a2 − 3b2) =
1
3(3− 2a2 − 2b2) = µr(g, pvs) (3.25)
as one can easily verify. Analogously, the amplitudes with the second and third basis vectors are
given by
〈0, 1, 0|vs〉 = aeiα〈0, 1, 0|vrb〉+ beiβ〈0, 1, 0|vry〉 =
√2
3bei(β+θy) (3.26)
〈0, 0, 1|vs〉 = aeiα〈0, 0, 1|vrb〉+ beiβ〈0, 0, 1|vry〉 =
√2
3aei(α+θb) (3.27)
respectively. Therefore, the transition probabilities are
|〈0, 1, 0|vs〉|2 =2
3b2 = µy(g, pvs) (3.28)
|〈0, 0, 1|vs〉|2 =2
3a2 = µb(g, pvs) (3.29)
From the foregoing follows that a general superposition state is represented by the unit vector
|vs〉 =1√3|aei(α+φr) + bei(β+θr),
√2bei(β+θy),
√2aei(α+θb)〉 (3.30)
and that the utilities corresponding to the different acts are given by
U(f1, g, pvs) = 〈vs|F1|vs〉 = 4$(3− 2a2 − 2b2) (3.31)
U(f2, g, pvs) = 〈vs|F2|vs〉 = 4$ · 2a2 (3.32)
U(f3, g, pvs) = 〈vs|F3|vs〉 = 4$ · (3− 2a2) (3.33)
U(f4, g, pvs) = 〈vs|F4|vs〉 = 4$(2a2 + 2b2) (3.34)
We can see that it is not necessarily the case that µr(g, pvs) = 13 , which means that choices of a
and b can be made such that the superposition state pvs /∈ ΣElls. The reason is that ΣElls is not a
linearly closed subset of C3.
Let us then consider some of the extreme possibilities of superpositions. We remind that the
latter superpositions are not relevant for the modeling of the Ellsberg paradox as it was originally
formulated, but they come into play if one wants to represent more general Ellsberg-type situations.
For example, choose a = b =√
32 . Then we have µy(g, pvs) = 2
3 · 34 = 1
2 , µb(g, pvs) = 23 · 3
4 = 12 , and
µr(g, pvs) = 0, and cos(β − α+ θr − φr) = 3(1−a2−b2)2ab = −1, hence β = π + α− θr + φr. Thus, the
state represented by
|vs(yb)〉 =
√3
2(eiα|vrb〉+ ei(π+α−θr+φr)|vry〉) (3.35)
gives rise to probability zero for a red ball to be drawn. Another extreme choice is, when we take
a = b =√
38 . Then we have µy(g, pvs) = 2
3 · 38 = 1
4 , µb(g, pvs) = 23 · 3
8 = 14 and µr(g, pvs) = 1
2 , and
CHAPTER 3. IDENTIFYING QUANTUM STRUCTURES IN THE ELLSBERG PARADOX32
cos(β − α+ θr − φr) = +1, which means β = α− θr + φr. Hence, the state
|vs(r)〉 =
√3
8(eiα|vrb〉+ ei(α−θr+φr)|vry〉) (3.36)
gives rise to probability 12 for a red ball to be drawn. These are extreme superposition states which
are not compatible with the situation as formulated by Ellsberg, but they could be useful if suitable
Ellsberg–type extensions are taken into account.
To construct non–trivial superpositions that retain probability 13 for drawing a red ball, we
require that1
3= µr(g, pvs) =
1
3(3− 2a2 − 2b2) (3.37)
or, equivalently,
a2 + b2 = 1 (3.38)
which implies that cos(β − α+ θr − φr) = 0, hence β = π2 + α− θr + φr.
Let us construct now two examples of superposition states that conserve the 13 probability for
drawing a red ball, and hence are conservative superpositions, and express ambiguity as is thought
to be the case in the Ellsberg paradox situation. The first state refers to the comparison for a bet
between f1 and f2. The ambiguity of not knowing the number of yellow and black balls in the
urn, only their sum to be 60, as compared to knowing the number of red balls in the urn to be
30, gives rise to the thought that ‘eventually there are perhaps almost no black balls and hence
an abundance of yellow balls’. Jointly, and in superposition, the thought also comes that ‘it is
of course also possible that there are more black balls than yellow balls’. These two thoughts in
superposition, are mathematically represented by a state pvs . The state pvs will be closer to pvry ,
the extreme state with no black balls, if the person is deeply ambiguity averse, while it will be closer
to pvrb , the extreme state with no yellow balls, if the person is attracted by ambiguity. Hence, these
two tendencies are expressed by the values of a and b in the superposition state. If we consider
again the utilities, this time with a2 + b2 = 1, we have
U(f1, g, pvs) = 4$ (3.39)
U(f2, g, pvs) = 4$ · 2a2 (3.40)
U(f3, g, pvs) = 4$ · (3− 2a2) (3.41)
U(f4, g, pvs) = 8$ (3.42)
So, for a2 < 12 , which exactly means that the superposition state pvs is closer to the state pvry
than to the state pvrb , we have that U(f2, g, pvs) < U(f1, g, pvs), and hence a person with strong
ambiguity aversion in the situation of the first bet, will then prefer to bet on f1 and not on f2. Let us
choose a concrete state for the bet between f1 and f2, and call it pv12s , and denote its superposition
state by |v12s 〉. Hence, for |v12
s 〉 we take a = 12 and b =
√3
2 and hence a2 = 14 and b2 = 3
4 . For the
angles we must have β−α+ θr − φr = π2 , hence let us choose θr = φr = 0, α = 0, and β = π
2 . This
CHAPTER 3. IDENTIFYING QUANTUM STRUCTURES IN THE ELLSBERG PARADOX33
gives us
|v12s 〉 =
1
2√
3|1 +
√3ei
π2 ,√
2√
3eiπ2 ,√
2〉 =1
2√
3|1 + i
√3, i√
6,√
2〉 (3.43)
On the other hand, for 12 < a2, which means that the superposition state is closer to the state pvrb
than to the state pvry , we have that U(f3, g, pvs) < U(f4, g, pvs), and hence a person with strong
ambiguity aversion in the situation of the second bet, will then prefer to bet on f4 and not on f3.
Also for this case we construct an explicit state, let us call it pv23s , and denote it by the vector |v34s 〉.
Hence, for |v34s 〉 we take a =
√3
2 and b = 12 and hence a2 = 3
4 and b2 = 14 . For the angles we must
have β − α+ θr − φr = π2 , hence let us choose θr = φr = 0, α = 0, and β = π
2 . This gives us
|v34s 〉 =
1
2√
3|√
3 + eiπ2 ,√
2eiπ2 ,√
2√
3〉 =1
2√
3|√
3 + i, i√
2,√
6〉 (3.44)
The superposition states p12vs and p34
vs representing the unit vectors |v12s 〉 and |v34
s 〉, respectively,
will be used in the next sections to provide a faithful modeling of a concrete experiment in which
decisions are expressed by real agents.
3.3 An experiment testing the Ellsberg paradox
We have observed in the previous sections that genuine quantum aspects intervene in the de-
scription of the Ellsberg paradox. This will be even more evident from the analysis of a statistically
relevant experiment we performed of this paradox, whose results were firstly reported in [14]. To
perform the experiment we sent out the following text to several people, consisting of a mixture of
friends, relatives and students, to avoid as much as possible a statistical selection bias.
We are conducting a small-scale statistics investigation into a particular problem and would like
to invite you to participate as test subjects. Please note that it is not the aim for this problem to
be resolved in terms of correct or incorrect answers. It is your preference for a particular choice
we want to test. The question concerns the following situation. Imagine an urn containing 90 balls
of three different colors: red balls, black balls and yellow balls. We know that the number of red
balls is 30 and that the sum of the the black balls and the yellow balls is 60. The questions of our
investigation are about the situation where somebody randomly takes one ball from the urn.
(i) The first question is about a choice to be made between two bets: bet f1 and bet f2. Bet f1
involves winning ‘10 euros when the ball is red’ and ‘zero euros when it is black or yellow’. Bet f2
involves winning ‘10 euros when the ball is black’ and ‘zero euros when it is red or yellow’. The
question we would ask you to answer is: Which of the two bets, bet f1 or bet f2, would you prefer?
(ii) The second question is again about a choice between two different bets, bet f3 and bet f4.
Bet f3 involves winning ‘10 euros when the ball is red or yellow’ and ‘zero euros when the ball is
black’. Bet f4 involves winning ‘10 euros when the ball is black or yellow’ and ‘zero euros when the
ball is red’. The second question therefore is: Which of the two bets, bet f3 or bet f4, would you
prefer?
Please provide in your reply message the following information.
CHAPTER 3. IDENTIFYING QUANTUM STRUCTURES IN THE ELLSBERG PARADOX34
For question 1, your preference (your choice between bet f1 and bet f2). For question 2, your
preference (your choice between bet f3 and bet f4). By ‘preference’ we mean ‘the bet you would take if
this situation happened to you in real life’. You are expected to choose one of the bets for each of the
questions, i.e. ‘not choosing is no option’. You are welcome to provide a brief explanation of your
preferences, which may be of a purely intuitive nature, only mentioning feelings, for example, but
this is not required. It is all right if you only state your preferences without giving any explanation.
One final remark about the colors. Your choices should not be affected by any personal color
preference. If you feel that the colors of the example somehow have an influence on your choices,
you should restate the problem and take colors that are indifferent to yours, if this does not work,
use other neutral characteristics to distinguish the balls.
Let us now analyze the obtained results.
We had 59 respondents participating in our test of the Ellsberg paradox problem, which is
the typical number of participants in experiments on psychological effect of the type studied by
Kahneman and Tversky, such as the conjunction fallacy, and the disjunction effect. (see, e.g.,
[29, 30]). We do believe that ambiguity aversion is a psychological effect within this calls of effects,
which means that in case our hypothesis on the nature of ambiguity aversion is correct, our test is
significant. This being said, it would certainly be interesting to make a similar test with a larger
number of participants, which is something we plan for the future. We however also remark that
the quantum modeling scheme is general enough to also model statistical data that are different
from the ones collected in this specific test. Next to this remark, we want to point out that in the
present paper we want to prove that these real data ‘can’ be modeled in our approach.
The answers of the participants were distributed as follows: (a) 34 subjects preferred bets f1
and f4; (b) 12 subjects preferred bets f2 and f3; (c) 7 subjects preferred bets f2 and f4; (d) 6
subjects preferred bets f1 and f3. This makes the weights with preference of bet f1 over bet f2 to
be 0.68 against 0.32, and the weights with preference of bet f4 over bet f3 to be 0.69 against 0.31.
It is worth to note that 34+12=46 people chose the combination of bet f1 and bet f4 or bet f2 and
bet f3, which is 78%. In Sect. 3.4 we apply our quantum mechanical model to these experimental
data.
3.4 Quantum modeling the experiment
As anticipated in Sect. 3.2.3, we take into account the superposition states p12vs and p34
vs to put
forward a description of the choices made by the participants in the test described in Sect. 3.3.
We employ in this section the spectral methods that are typically used in quantum mechanics to
construct self–adjoint operators.
First we consider the choice to bet on f1 or on f2. This is a choice with two possible outcomes,
let us call them o1 and o2. Thus, we introduce two projection operators P1 and P2 on the Hilbert
space C3, and represent the observable associated with the first bet by the self–adjoint operator
(spectral decomposition) O12 = o1P1 + o2P2. Then, we consider the choice to bet on f3 or on f4.
This is a choice with two possible outcomes too, let us call them o3 and o4. Thus, we introduce two
CHAPTER 3. IDENTIFYING QUANTUM STRUCTURES IN THE ELLSBERG PARADOX35
projection operators P3 and P4 on C3, and represent the observable associated with the second bet
by the self–adjoint operator (spectral decomposition) O34 = o3P3 + o4P4.
To model the empirical data in Sect. 3.3, we recall that we tested in our experiment 59 partic-
ipants, and 40 preferred f1 over f2, while the remaining 19 preferred f2 over f1. This means that
we should construct P1 and P2 in such a way that
〈v12s |P1|v12
s 〉 =40
59= 0.68 〈v12
s |P2|v12s 〉 =
19
59= 0.32 (3.45)
In our experiment of the 59 participants there were 41 preferring f4 over f3, and 18 who choose the
other way around. Hence, we should have
〈v34s |P3|v34
s 〉 =18
59= 0.31 〈v34
s |P4|v34s 〉 =
41
59= 0.69 (3.46)
Both bets should give rise to no preference, hence probabilities 12 in all cases, when there is no
ambiguity, when the state is pvc represented by the vector
|vc〉 = | 1√3,
1√3,
1√3〉 (3.47)
Let us preliminarily denote by S121 , S12
2 , S343 and S34
4 the eigenspaces associated with the eigenvalues
o1, o2, o3 and o4, respectively. We can assume that S122 and S34
4 are one dimensional, and S121 and
S343 are two dimensional, without loss of generality. Then, we prove the following two theorems.
Theorem 1. If, under the hypothesis on the eigenspaces formulated above, the two self–adjoint
operators (spectral decompositions) O12 = o1P1 + o2P2 and O34 = o3P3 + o4P4 are such that
[O12,O34] = 0 in C3, then two situations are possible, (i) an orthonormal basis {|e1〉, |e2〉, |e3〉} of
common eigenvectors exists such that P1 = |e1〉〈e1|+ |e2〉〈e2|, P2 = |e3〉〈e3|, P3 = |e2〉〈e2|+ |e3〉〈e3|and P4 = |e1〉〈e1|, or (ii) O12 = O34, and hence S12
2 = S344 and S12
1 = S343 .
Proof: Suppose that O12 6= O34. Since S122 and S34
4 are one dimensional, we can choose |e1〉 and
|e3〉 unit vectors respectively in S122 and S34
4 . Since [O12,O34] = 0 it follows that [P2, P4] = 0, and
hence |e1〉 ⊥ |e3〉 or |e1〉 = λ|e2〉 for some λ ∈ C. But, if |e1〉 = λ|e2〉, we have S122 = S34
4 , and hence
S121 = (S12
2 )⊥ = (S344 )⊥ = S34
3 , which entails O12 = O34. Hence, we have |e1〉 ⊥ |e3〉. Since S121 and
S343 are both two dimensional, their intersection is one dimensional, or they are equal. If S12
1 = S343 ,
then also (S121 )⊥ = (S34
3 )⊥, which would entail again that O12 = O34. Hence, we have that their
intersection is one dimensional. We choose |e2〉 a unit vector contained in this intersection, and
hence |e2〉 ⊥ |e1〉 and |e2〉 ⊥ |e3〉. This means that {|e1〉, |e2〉, |e3〉} is an orthonormal basis, and
P1 = |e1〉〈e1|+ |e2〉〈e2|, P2 = |e3〉〈e3|, P3 = |e2〉〈e2|+ |e3〉〈e3| and P4 = |e1〉〈e1|.
Theorem 2. For the data of our experiment exist compatible self-adjoint operators, hence, following
the notations introduced, O12 6= O34 such that [O12,O34] = 0, modeling both bets. This means that,
CHAPTER 3. IDENTIFYING QUANTUM STRUCTURES IN THE ELLSBERG PARADOX36
again following the notations introduced, that we have
〈e1|e1〉 = 〈e2|e2〉 = 〈e3|e3〉 = 1 (3.48)
〈e1|e2〉 = 〈e1|e3〉 = 〈e2|e3〉 = 0 (3.49)
|〈v12s |e1〉|2 + |〈v12
s |e2〉|2 = 0.68 (3.50)
|〈v12s |e3〉|2 = 0.32 (3.51)
|〈v34s |e2〉|2 + |〈v34
s |e3〉|2 = 0.31 (3.52)
|〈v34s |e1〉|2 = 0.69 (3.53)
|〈vc|e1〉|2 + |〈vc|e2〉|2 = 0.5 = |〈vc|e3〉|2 (3.54)
|〈vc|e2〉|2 + |〈vc|e3〉|2 = 0.5 = |〈vc|e1〉|2 (3.55)
Proof: We can explicitly construct an orthonormal basis {|e1〉, |e2〉, |e3〉} which simultaneously
satisfies Eqs. (3.48)–(3.55). We omit the explicit construction, for the sake of brevity, and we only
report the solution, as follows. The orthonormal vectors |e1〉 and |e3〉 are respectively given by
|e1〉 = |0.38ei61.2◦ , 0.13ei248.4◦ , 0.92ei194.4◦〉 (3.56)
|e3〉 = |0.25ei251.27◦ , 0.55ei246.85◦ , 0.90ei218.83◦〉 (3.57)
One can then construct at once a unit vector |e2〉 orthogonal to both |e1〉 and |e3〉, which we don’t
do explicitly, again for the sake of brevity.
Theorems 1 and 2 entail that within our quantum modeling approach the two bets can be
represented by commuting observables such that the statistical data of the real experiment are
faithfully modeled.
The following orthogonal projection operators then model the agents’ decisions.
P2 = |e3〉〈e3| =
0.06 0.14ei4.42◦ 0.23ei32.44◦
0.14e−i4.42◦ 0.30 0.49ei28.02◦
0.23e−i32.44◦ 0.49e−i28.02◦ 0.81
(3.58)
and
P4 = |e1〉〈e1| =
0.14 0.05e−i187.2◦ 0.35e−i133.2◦
0.05ei187.2◦ 0.02 0.12ei54◦
0.35ei133.2◦ 0.12e−i54◦ 0.85
(3.59)
Thus, P1 = 1− P2 and P3 = 1− P4 can be easily calculated.
Let us now come to the representation of the observables. The observable associated with the
preference between f1 and f2 is then represented by the self–adjoint operator (spectral decomposi-
tion) O12, while the observable associated with the preference between f3 and f4 is represented by
the self–adjoint operator (spectral decomposition) O34. More explicitly, if we set o1 = o3 = 1 and
CHAPTER 3. IDENTIFYING QUANTUM STRUCTURES IN THE ELLSBERG PARADOX37
o2 = o4 = −1, we get the following explicit representations.
O12 = P1 − P2 = 1− 2P2
=
0.87 −0.28ei4.42◦ −0.46ei32.44◦
−0.28e−i4.42◦ 0.40 −0.98ei28.02◦
−0.45e−i32.44◦ −0.98e−i28.02◦ −0.62
(3.60)
O34 = P3 − P4 = 1− 2P4
=
0.71 −0.10e−i187.2◦ −0.70e−i133.2◦
−0.10ei187.2◦ 0.97 −0.24ei54◦
−0.70ei133.2◦ −0.24e−i54◦ −0.69
(3.61)
A direct calculation of the commutator operator [O12,O34] reveals that the corresponding observ-
ables are indeed compatible.
3.5 A real vector space analysis
We show in this section that the possibility of representing the experimental data in Sect. 3.3
by compatible measurements for the bets relies crucially on our choice of a Hilbert space over
complex numbers as a modeling space. Indeed, if a Hilbert space over real numbers is attempted,
no compatible observables for the bets and the data in Sect. 3.3 can be constructed any longer, as
the following analysis reveals.
Let us indeed attempt to represent the Ellsberg paradox situation in the real Hilbert space R3.
This comes to allowing only values of 0 and π for the phases of the complex vectors in Sect. 3.2.3.
It is then easy to see that the only conservative superpositions that remain are the extreme states
themselves, that is,
|v12s 〉 = |vry〉 = | ± 1√
3,±√
2
3, 0〉 (3.62)
|v34s 〉 = |vrb〉 = | ± 1√
3, 0,±
√2
3〉 (3.63)
This means that only superposition states such as |v12s 〉 and |v34
s 〉 are only conservative, in case
complex numbers are used for the superposition. This is a first instance of the necessity of complex
numbers for a quantum modeling of the Ellsberg situation, because indeed, we should be able to
represent the ‘priors’, hence the quantum states, by superpositions of the extreme states, not by the
extreme states themselves. But, let us prove that, even if we opt for representing the quantum states
by the extreme states, no real Hilbert space representation with compatible observables modeling
the bets and the experimental data is possible.
Taking into account the content of Th. 1, and making use of the notations introduced in Sect.
3.4, we can state the following: For a compatible solution to exist, we need to find two unit vectors
CHAPTER 3. IDENTIFYING QUANTUM STRUCTURES IN THE ELLSBERG PARADOX38
|x〉 and |y〉 such that |x〉 ∈ S122 and |y〉 ∈ S34
4 . For case (i) of Th. 1 to be satisfied, this is the case
where the self-adjoint operators representing the compatible measurements are different, we have
that |x〉 needs to be orthogonal to |y〉, which can be expressed as 〈x|y〉 = 0. For the case (ii) of
Th. 1 to be satisfied, this is the case where the self-adjoint operators representing the compatible
measurements are equal, we have that |x〉 needs to be a multiple of |y〉, and since both are unit
vectors, and we work in a real Hilbert space, this means that |x〉 = ±|y〉. This can be expressed as
〈x|y〉 = ±1. In the following we will prove that such |x〉 and |y〉 do not exist in a real Hilbert space.
In our proof we start by supposing that these vectors exist and find a contradiction. Let us put
|x〉 = |a, b, c〉 and |y〉 = |d, f, g〉. Since |x〉 and |y〉 are unit vectors, we have 1 = a2 + b2 + c2 =
d2 + f2 + g2. Generalizing Eqs. (3.51), (3.53), (3.54) and (3.55) to considering the two cases (i)
and (ii) of Th. 1, we have that the vectors |x〉 and |y〉 must satisfy following equations.
1
2= |〈 1√
3,
1√3,
1√3|a, b, c〉|2
=1
3(a+ b+ c)2 =
1
3(a2 + b2 + c2 + 2ab+ 2ac+ 2bc) (3.64)
0.69 = |〈 1√3, 0,
√2
3|a, b, c〉|2
=1
3(a+
√2c)2 =
1
3(a2 + 2c2 + 2
√2ac) (3.65)
1
2= |〈 1√
3,
1√3,
1√3|d, f, g〉|2
=1
3(d+ f + g)2 =
1
3(d2 + f2 + g2 + 2df + 2dg + 2fg) (3.66)
0.32 = |〈 1√3,
√2
3, 0|d, f, g〉|2
=1
3(d+
√2f)2 =
1
3(d2 + 2f2 + 2
√2df) (3.67)
We stress that we have considered here the ++ signs choices for |v12s 〉 and |v34
s 〉. We will later
consider the other possibilities. Elaborating we get the following set of equations to be satisfied
1 = a2 + b2 + c2 (3.68)
1.5 = a2 + b2 + c2 + 2ab+ 2ac+ 2bc (3.69)
2.07 = a2 + 2c2 + 2√
2ac (3.70)
1 = d2 + f2 + g2 (3.71)
1.5 = d2 + f2 + g2 + 2df + 2dg + 2fg (3.72)
0.96 = d2 + 2f2 + 2√
2df (3.73)
The points (a, b, c) satisfying Eq. (3.69) lie on a cone in the origin and centred around |vc〉, and the
points (a, b, c) satisfying Eq. (3.70) lie on a cone in the origin and centred around |v34s 〉. Hence Eqs.
(3.69) and (3.70) can only jointly be satisfied where these two cones intersect. Further need (a, b, c)
CHAPTER 3. IDENTIFYING QUANTUM STRUCTURES IN THE ELLSBERG PARADOX39
to be the coordinates of a unit vector, which is expressed by Eq. (3.68). For two cones there are
in a three dimensional real space only two possibilities, or they cut each other in two lines through
the origin, or they do not have an intersection different from the origin. On two lines through the
origin, four unit vectors can be found always. This means that Eqs. (3.68), (3.69) and (3.70) have
four solutions, or none. We are in a situation here of four solutions, which are the following
1. a = 0.052 b = 0.192 c = 0.980 (3.74)
2. a = −0.931 b = 0.065 c = −0.359 (3.75)
3. a = −0.052 b = −0.192 c = −0.980 (3.76)
4. a = 0.931 b = −0.065 c = 0.359 (3.77)
Also the solutions of Eqs. (3.71), (3.72) and (3.73) are four points on the two intersecting lines of
a cone, or none, if the cones do not intersect. In the situation corresponding to our experimental
data, we also here find four solutions, which are
1. d = 0.969 f = 0.008 g = 0.248 (3.78)
2. d = 0.154 f = −0.802 g = −0.577 (3.79)
3. d = −0.969 f = −0.008 g = −0.248 (3.80)
4. d = −0.154 f = 0.802 g = 0.577 (3.81)
Our proof follows now easily, when we verify that none of these solutions allows |x〉 and |y〉 to be
an orthogonal pair of vectors, or a pair of vectors equal to each other, or to ones opposite. We can
verify this by calculating the number 〈x|y〉, and seeing that they are all different from 0, different
from +1, and different from -1. We indeed have
〈e1|e3〉11 = 0.296 〈e1|e3〉12 = −0.711 (3.82)
〈e1|e3〉13 = −0.296 〈e1|e4〉13 = 0.711 (3.83)
〈e1|e3〉21 = −0.990 〈e1|e3〉22 = 0.011 (3.84)
〈e1|e3〉23 = 0.990 〈e1|e3〉24 = −0.011 (3.85)
〈e1|e3〉31 = −0.296 〈e1|e3〉32 = 0.711 (3.86)
〈e1|e3〉33 = 0.296 〈e1|e4〉33 = −0.711 (3.87)
〈e1|e3〉41 = 0.990 〈e1|e3〉42 = −0.011 (3.88)
〈e1|e3〉43 = −0.990 〈e1|e3〉44 = 0.011 (3.89)
To complete our proof of the non-existence of compatible observables in a real Hilbert space, we
need to analyse also all the other possibilities, i.e. all possible choices of + and − for |v12s 〉 and |v34
s 〉.This can be done completely along the same lines of the above, and hence we do not represent it
explicitly here. We have however verified all cases carefully, and indeed, none of the possibilities
CHAPTER 3. IDENTIFYING QUANTUM STRUCTURES IN THE ELLSBERG PARADOX40
lead to vectors |x〉 and |y〉 such that their inner product 〈x|y〉 equals 0, +1, or −1.
The above result is relevant, in our opinion, and we think it is worth to discuss it more in
detail. The existence of compatible observables to represent the decision-makers’ choice among the
different acts in our experiment on the Ellsberg paradox is a direct consequence of the fact that
we used a complex Hilbert space as a modeling space. As we have proved here, if one instead
uses a real vector space, then the collected data cannot be reproduced by compatible observables.
Hence, one has two possibilities, in this case. Either one requires that compatible observables exist
that accord with an Ellsberg-type situation, and then one has to accept a complex Hilbert space
representation where ambiguity aversion is coded into superposed quantum states (note that these
superpositions are of the ‘complex-type’, hence entailing genuine interference – superpositions with
complex (non-real) coefficients always entail that the quantum effect of interference is present).
Alternatively, one can use a representation in a real vector space but, then, one should accept
that an Ellserg-type situation cannot be reproduced by compatible observables. In either case, the
appearance of quantum structures – interference due to the presence of genuine complex numbers,
or incompatibility due to the impossibility to represent the data by compatible measurements –
seems unavoidable in the Ellsberg paradox situation.
Our quantum theoretic modeling of the Ellsberg paradox situation is thus completed. We
however want to add some explanatory remarks concerning the novelty of our approach, as follows.
(i) We have incorporated the subjective preference of traditional economics approaches in the
quantum state describing the conceptual Ellsberg entity. At variance with existing proposals, the
subjective preference coded in the quantum state can be different for each one of the acts fj , since it
is not derived from any prefixed mathematical rule. Therefore, we can naturally explain a situation
with f1 preferred to f2 and f4 preferred to f3, without the need of assuming extra hypotheses.
(ii) In the present paper, we have detected genuine quantum aspects in the description of
the Ellsberg paradox situation, namely, contextuality, superposition, and the ensuing interference.
Further deepening and experimenting most probably will reveal other quantum aspects. Hence the
hypothesis that quantum effects, of a conceptual nature, concretely drive decision–makers’ behavior
in uncertainty situations is warranted.
(iii) We have focused here on the Ellsberg paradox. But, our quantum theoretic modeling is
sufficiently general to cope with various generalizations of the Ellsberg paradox, which are problem-
atical in traditional economics approaches, such as the Machina paradox [17]. This opens the way
toward the construction of a unified framework extending standard expected utility and modeling
‘ambiguity laden’ situations in economics and decision theory.
Bibliography
[1] von Neumann, J., Morgenstern, O.: Theory of Games and Economic Behavior. Princeton
University Press, Princeton (1944)
[2] Savage, L.J.: The Foundations of Statistics. Wiley, New York (1954)
[3] Ellsberg, D.: Risk, ambiguity, and the Savage axioms. Quart. J. Econ. 75, 643–669 (1961)
[4] Machina, M.J.: Risk, Ambiguity, and the Dark–dependence Axioms. Am. Econ. Rev. 99,
385–392 (2009)
[5] Knight, F.H.: Risk, Uncertainty and Profit. Houghton Mifflin, Boston (1921)
[6] I. Gilboa, Expected utility with purely subjective non-additive probabilities, J. Math. Econ.
16, 65–88 (1987)
[7] Gilboa, I., Schmeidler, D.: Maxmin expected utility with non–unique prior. J. Math. Econ.
18, 141–153 (1989)
[8] Maccheroni, F., Marinacci, M., Rustichini, A.: Dynamical variational preferences. The Carlo
Alberto Notebooks 1, pp. 37 (2006)
[9] Klibanoff, P., Marinacci, M., Mukerji, S.: A smooth model of decision making under ambiguity.
Econometrica 73, 1849–1892 (2005)
[10] Epstein, L.G.: A Definition of uncertainty aversion. Rev. Econ. Stud. 66, 579–608 (1999)
[11] Baillon, A., L’Haridon, O., Placido, L.: Ambiguity models and the Machina paradoxes. Am.
Econ. Rev. 1547, 1547–1560 (2011)
[12] Gudder, S.P.: Quantum probability spaces. Proc. Amer. Math. Soc. 21, 296302 (1969)
[13] Aerts, D., Broekaert, J., Czachor, M., D’Hooghe, B.: The violation of expected utility hy-
pothesis in the disjunction effect. A quantum-conceptual explanation of violations of expected
utility in economics. Lecture Notes in Computer Science vol. 7052, 192–198. Springer, Berlin
(2011)
[14] Aerts, D., D’Hooghe, B., Sozzo, S.: A quantum cognition analysis of the Ellsberg paradox.
Lecture Notes in Computer Science vol. 7052, 95–104. Springer, Berlin (2011)
41
BIBLIOGRAPHY 42
[15] Aerts, D., Sozzo, S.: Quantum Structure in Economics: The Ellsberg Paradox. In: D’Ariano,
M., et al. (eds.) Quantum Theory: Reconsideration of Foundations–6, pp. 487–494. AIP, New
York (2012)
[16] Aerts, D., Sozzo, S.: Contextual Risk and Its Relevance in Economics; A Contextual Risk
Model for the Ellsberg Paradox. J. Eng. Sci. Tech. Rev. 4, 241-245; 246–250 (2012)
[17] Aerts, D., Sozzo, S., Tapia, J.: A quantum model for the Ellsberg and Machina paradoxes.
Lecture Notes in Computer Science vol. 7620, 48–59. Springer, Berlin (2012)
[18] Bruza, P.D., Lawless, W., van Rijsbergen, C.J., Sofge, D., Editors: Proceedings of the AAAI
Spring Symposium on Quantum Interaction, March 27–29. Stanford University, Stanford
(2007)
[19] Bruza, P.D., Lawless, W., van Rijsbergen, C.J., Sofge, D., Editors: Quantum Interaction:
Proceedings of the Second Quantum Interaction Symposium. College Publications, London
(2008)
[20] Bruza, P.D., Sofge, D., Lawless, W., Van Rijsbergen, K., Klusch, M., Editors: Proceedings of
the Third Quantum Interaction Symposium. Lecture Notes in Artificial Intelligence vol. 5494.
Springer, Berlin (2009)
[21] Aerts, D.: Quantum structure in cognition. J. Math. Psychol. 53, 314–348 (2009)
[22] Busemeyer, J.R., Lambert-Mogiliansky, A.: An exploration of type indeterminacy in strategic
decision-making. Lecture Notes in Computer Science vol. 5494, 113–127. Springer, Berlin
(2009)
[23] Pothos, E.M., Busemeyer, J.R.: A quantum probability model explanation for violations of
‘rational’ decision theory. Proc. Roy. Soc. B 276, 2171–2178 (2009)
[24] Danilov, V.I., Lambert-Mogiliansky, A.: Expected utility theory under non-classical uncer-
tainty. Theory and Decision 68, 25–47 (2010)
[25] Khrennikov, A.Y.: Ubiquitous Quantum Structure. Springer, Berlin (2010)
[26] Song, D., Melucci, M., Frommholz, I., Zhang, P., Wang, L., Arafat, S., Editors: Quantum
Interaction. Lecture Notes in Computer Science vol. 7052. Springer, Berlin (2011)
[27] Busemeyer, J.R., Pothos, E., Franco, R., Trueblood, J.S.: A quantum theoretical explanation
for probability judgment ‘errors’. Psychol. Rev. 118, 193–218 (2011)
[28] Busemeyer, J.R., Bruza, P.D.: Quantum Models of Cognition and Decision. Cambridge Uni-
versity Press, Cambridge (2012)
BIBLIOGRAPHY 43
[29] Tversky, A., Kahneman, D.: Judgements of and by representativeness. In: Kahneman, D.,
Slovic, P., Tversky, A. (Eds.) Judgement Under Uncertainty: Heuristics and Biases, 84–98.
Cambridge University Press, Cambridge (1982)
[30] Tversky, A., Shafir, E.: The disjunction effect in choice under uncertainty. Psychol. Sci. 3,
305–309 (1992)
Appendix A
Appendix Welfare Cost of Model
Uncertainty in a SOE
A.1 The model’s first-order conditions
In this section I report the first-order conditions that come from household’s maximizationproblem. Assuming that households have a GHH Bernoulli utility function and preferences forrobustness.
(ct −
hωtω
)−σ= β(1 + rt)Et
{exp(
−Vt+1
θ)
Et exp(−Vt+1
θ)
(ct+1 −
hωt+1
ω
)−σ}(A.1)
(ct −
hωtω
)−σ(1 + φ(kt+1 − kt)) = βEt
{exp(
−Vt+1
θ)
Et exp(−Vt+1
θ)
(ct+1 −
hωt+1
ω
)−σ [αeat+1kα−1
t+1 h1−αt+1 + 1− δ + φ(kt+2 − kt+1)
]}(A.2)
hω−1t = (1− α)eatkαt h
−αt (A.3)
A.1.1 Model’s deterministic steady state solution
In the deterministic steady state, the productivity shock is constant and equal to one and
the other state variables do not change over time. The equations characterizing the steady state
comprise the definition of the value function (A.4), the first-order conditions (A.5)-(A.7) and the
budget constraint(A.8):
Vss =
(css − hωss
ω
)1−σ− 1
(1− σ)(1− β)(A.4)
1 = β(1 + r∗) (A.5)
1 = β(αeakα−1ss h1−α
ss + (1− δ)) (A.6)
hω−1ss = (1− α)ea
(k
h
)α(A.7)
css = yss − r∗d− iss (A.8)
44
APPENDIX A. APPENDIX WELFARE COST OF MODEL UNCERTAINTY IN A SOE 45
Then using (A.6):
ksshss
=
[1β − (1− δ)
αea
] 1α−1
= κ (A.9)
From the steady-state labor supply (A.7), I obtain:
hss = [(1− α)eaκα]1
ω−1 (A.10)
kss = κhss (A.11)
(A.12)
A.2 Model solution
Here, I present the first-order derivatives of the value function with respect to the state variables,
calculated using the envelope theorem with restrictions, in order to find the first-order coefficients
of the Taylor approximation:
Vd,t = −λt[(1 + r∗t + ψ(edt−d − 1)) + ψdtedt−d] (A.13)
Vk,t = λt[(1− δ) + αatkα−1t h1−α
t + φ(kt+1 − kt)] (A.14)
Va,t = βEt
{exp(−Vt+1
θ )ρVa,t+1
}
Et exp(−Vt+1
θ )+ λtk
αt h
1−αt (A.15)
Vχ,t = βEt
{exp(−Vt+1
θ )(Va,t+1(ηεt+1) + Vχ,t+1)}
Et exp(−Vt+1
θ )(A.16)
Evaluating at the deterministic steady state:
Vd,ss = −λss[(1 + r∗t ) + ψdss] (A.17)
Vk,ss = λss[(1− δ) + αeasskα−1ss h1−α
ss ] (A.18)
Va,ss =λssk
αssh
1−αss
1− βρ (A.19)
Vχ,ss = 0 (A.20)
The second-order derivatives of the value function with respect to the state varibles are:
APPENDIX A. APPENDIX WELFARE COST OF MODEL UNCERTAINTY IN A SOE 46
Vχχ,t = β
Et
{exp(
−Vt+1θ
) · − 1θ(Vd,t+1dχ,t+1 + Vk,t+1kχ,t+1 + Va,t+1ηεt+1 + Vχ,t+1)(Va,t+1ηεt+1 + Vχ,t+1)
}
Et exp(−Vt+1θ
)(A.21)
+β
Et
{exp(
−Vt+1θ
)(Va,t+1ηεt+1 + Vχ,t+1)
}Et
{exp(
−Vt+1θ
) · 1θ(Vd,t+1dχ,t+1 + Vk,t+1kχ,t+1 + Va,t+1ηεt+1 + Vχ,t+1)
}
[Et exp(−Vt+1θ
)]2
+β
Et
{exp(
−Vt+1θ
)((Vad,t+1dχ,t+1 + Vak,t+1kχ,t+1 + Vaa,t+1ηεt+1 + Vaχ,t+1)ηεt+1 + Vχd,t+1dχ,t+1 + Vχk,t+1kχ,t+1 + Vχa,t+1cεt+1 + Vχχ,t+1)
}
Et exp(−Vt+1θ
)
Evaluating at the deterministic steady state:
Vχχ,ss = −βθV 2a,ssη
2 + βVaa,ssη2 + βVχχ,ss (A.22)
⇒ Vχχ,ss =βη2
1− β
[Vaa,ss −
V 2a,ss
θ
](A.23)
The value function second-order derivatives and equilibrium conditions first-order derivatives.
dd,t = 1 + rf + dtedt−df ψ +
(−1 + e
dt−df)ψ + cd,t − ath−α
t kαt (1− α)hd,t + kd,t+1 + (−kt + kt+1)φkd,t+1 (A.24)
dk,t+1 = −1 + δ + ct,k − at(h1−αt k
−1+αt α + h
−αt k
αt (1− α)ht,k
)+ (−kt + kt+1)φ
(−1 + kk,t+1
)+ kk,t+1 (A.25)
da,t+1 = −h1−αt k
αt + ct,a − ath−α
t kαt (1− α)ht,a + ka,t+1 + (−kt + kt+1)φka,t+1 (A.26)
dχ,t+1 = cχ,t − ath−αt k
αt (1− α)hχ,t + kχ,t+1 + (−kt + kt+1)φkχ,t+1 (A.27)
− σ(ct −
hωt
ω
)−1−σ (cd,t − h−1+ω
t hd,t
)= e
dt−dfβψE
e−Vt+1θ
(ct+1 −
hωt+1ω
)−σ
E
[e−Vt+1θ
]
+ β(1 + rf +
(−1 + e
dt−df)ψ)
E
e−Vt+1θ
(ct+1 −
hωt+1ω
)−σE
e−Vt+1θ
(Va,t+1ad,t+1+Vk,t+1kd,t+1+dd,t+1Vd,t+1
)
θ
E
[e−Vt+1θ
]2
−σ
(ct+1 −
hωt+1ω
)−1−σ (cd,t+1 − h−1+ω
t+1 hd,t+1
)
E
[e−Vt+1θ
]
−E
e−Vt+1θ
(ct+1 −
hωt+1ω
)−σ (Va,t+1ad,t+1 + Vk,t+1kd,t+1 + dd,t+1Vd,t+1
)
θE
[e−Vt+1θ
]
(A.28)
APPENDIX A. APPENDIX WELFARE COST OF MODEL UNCERTAINTY IN A SOE 47
− σ(c−
hωt
ω
)−1−σ (ck,t − h−1+ω
hk,t
)= β
(1 + rf +
(−1 + e
dt−df)ψ)
E
e−Vt+1θ
(ct+1 −
hωt+1ω
)−σE
e−Vt+1θ
(Va,t+1ak,t+1+kt+1Vk,t+1+dk,t+1Vd,t+1
)
θ
E
[e−Vt+1θ
]2
−σ
(ct+1 −
hωt+1ω
)−1−σ (ck,t+1 − h−1+ω
t+1 hk,t+1
)
E
[e−Vt+1θ
]
−E
e−Vt+1θ
(ct+1 −
hωt+1ω
)−σ (Va,t+1ak,t+1 + kk,t+1Vk,t+1 + dk,t+1Vd,t+1
)
θE
[e−Vt+1θ
]
(A.29)
− σ(ct −
hωt
ω
)−1−σ (ca,t − h−1+ω
t ha,t
)= β
(1 + rf +
(−1 + e
dt−df)ψ)
E
e−Vt+1θ
(ct+1 −
hωt+1ω
)−σE
e−Vt+1θ
(aa,t+1Va,t+1+ka,t+1Vk,t+1+da,t+1Vd,t+1
)
θ
E
[e−Vt+1θ
]2
−σ
(ct+1 −
hωt+1ω
)−1−σ (ca,t+1 − h−1+ω
t+1 ha,t+1
)
E
[e−Vt+1θ
]
−E
e−Vt+1θ
(ct+1 −
hωt+1ω
)−σ (aa,t+1Va,t+1 + ka,t+1Vk,t+1 + da,t+1Vd,t
)
θE
[e−Vt+1θ
]
(A.30)
− σ(ct −
hωt
ω
)−1−σ (ct,χ − h−1+ω
t hχ,t
)= β
(1 + rf +
(−1 + e
dt−df)ψ)
E
e−Vt+1θ
(ct+1 −
hωt+1ω
)−σE
e−Vt+1θ
(Vχ,t+1+aχ,t+1Va,t+1+kχt+1Vk,t+1+dχ,t+1Vd,t+1
)
θ
E
[e−Vt+1θ
]2
−σ
(ct+1 −
hωt+1ω
)−1−σ (cχ,t+1 − h−1+ω
t+1 hχ,t+1
)
E
[e−Vt+1θ
]
−E
e−Vt+1θ
(ct+1 −
hωt+1ω
)−σ (Vχ,t+1 + aχ,t+1Va,t+1 + kχ,t+1Vk,t+1 + dχ,t+1Vd,t+1
)
θE
[e−Vt+1θ
]
(A.31)
APPENDIX A. APPENDIX WELFARE COST OF MODEL UNCERTAINTY IN A SOE 48
− σ(1 + (−kt + kt+1)φ)
(ct −
hωt
ω
)−1−σ (cd,t − h−1+ω
t hd,t
)+ φ
(ct −
hωt
ω
)−σkd,t+1 = β
(E
[e−Vt+1θ
(1 + at+1h
1−αt+1 k
−1+αt+1 α− δ + (−kt+1 + kt+2)φ
)(ct+1 −
hωt+1ω
)−σE
e−Vt+1θ
(Va,t+1ad,t+1+Vk,t+1kd,t+1+dd,t+1Vd,t+1
)
θ
E
[e−Vt+1θ
]2
+1
E
[e−Vt+1θ
](−σ
(1 + at+1h
1−αt+1 k
−1+αt+1 α− δ + (−kt+1 + kt+2)φ
)(ct+1 −
hωt+1
ω
)−1−σ (cd,t+1 − h−1+ω
t+1 hd,t+1
)
+
(ct+1 −
hωt+1
ω
)−σ (α(h1−αt+1 k
−1+αt+1 ad,t+1 + at+1
(h−αt+1k
−1+αt+1 (1− α)hd,t+1 + h
1−αt+1 k
−2+αt+1 (−1 + α)kd,t+1
))
+φ(ka,t+2ad,t+1 − kd,t+1 + kk,t+2kd,t+1 + dd,t+1kd,t+2
))))]
−E
e−Vt+1θ
(1 + at+1h
1−αt+1 k
−1+αt+1 α− δ + (−kt+1 + kt+2)φ
)(ct+1 −
hωt+1ω
)−σ (Va,t+1ad,t+1 + Vk,t+1kd,t+1 + dd,t+1Vd,t+1
)
θE
[e−Vt+1θ
]
(A.32)
− σ(1 + (−kt + kt+1)φ)
(ct −
hωt
ω
)−1−σ (ck,t − h−1+ω
t hk,t
)+ φ
(ct −
hωt
ω
)−σ (−1 + kk,t+1
)= β
(E
[e−Vt+1θ
(1 + at+1h
1−αt+1 k
−1+αt+1 α− δ + (−kt+1 + kt+2)φ
)(ct+1 −
hωt+1ω
)−σE
e−Vt+1θ
(Va,t+1ak,t+1+kk,t+1Vk,t+1+dk,t+1Vd,t+1
)
θ
E
[e−Vt+1θ
]2
+1
E
[e−Vθ
](−σ
(1 + at+1h
1−αt+1 k
−1+αt+1 α− δ + (−kt+1 + kt+2)φ
)(ct+1 −
hωt+1
ω
)−1−σ (ck,t+1 − h−1+ω
t+1 hk,t+1
)
+
(ct+1 −
hωt+1
ω
)−σ (α(h1−αt+1 k
−1+αt+1 ak,t+1 + at+1
(h−αt+1k
−1+αt+1 (1− α)hk,t+1 + h
1−αt+1 k
−2+αt+1 (−1 + α)kk,t+1
))
+φ(ka,t+2ak,t+1 − kk,t+1 + kk,t+1kk,t+2 + dk,t+1kd,t+2
))))]
−E
e−Vt+1
θ(1 + at+1h
1−αt+1 k
−1+αt+1 α− δ + (−kt+1 + kt+2)φ
)(ct+1 −
hωt+1ω
)−σ (Va,t+1ak,t+1 + kk,t+1Vk,t+1 + dk,t+1Vd,t+1
)
θE
[e−Vt+1θ
]
(A.33)
APPENDIX A. APPENDIX WELFARE COST OF MODEL UNCERTAINTY IN A SOE 49
− σ(1 + (−kt + kt+1)φ)
(ct −
hωt
ω
)−1−σ (ca,t − h−1+ω
ha,t
)+ φ
(ct −
hωt
ω
)−σka,t+1 = β
(E
[e−Vt+1θ
(1 + at+1h
1−αt+1 k
−1+αt+1 α− δ + (−kt+1 + kt+2)φ
)(ct+1 −
hωt+1ω
)−σE
e−Vt+1θ
(aa,t+1Va,t+1+kd,t+1Vk,t+1+da,t+1Vd,t+1
)
θ
E
[e−Vt+1θ
]2
+1
E
[e−Vt+1θ
](−σ
(1 + at+1h
1−αt+1 k
−1+αt+1 α− δ + (−kt+1 + kt+2)φ
)(ct+1 −
hωt+1
ω
)−1−σ (ca,t+1 − h−1+ω
t+1 ha,t+1
)
+
(ct+1 −
hωt+1
ω
)−σ (α(h1−αt+1 k
−1+αt+1 aa,t+1 + at+1
(h−αt+1k
−1+αt+1 (1− α)ha,t+1 + h
1−αt+1 k
−2+αt+1 (−1 + α)ka,t+1
))
+ φ(−ka,t+1 + aa,t+1ka,t+2 + ka,t+1kk,t+2 + da,t+1kd,t+2
))))]
−E
e−Vt+1θ
(1 + at+1h
1−αt+1 k
−1+αt+1 α− δ + (−kt+1 + kt+2)φ
)(ct+1 −
hωt+1ω
)−σ (aa,t+1Va,t+1 + ka,t+1Vk,t+1 + da,t+1Va,t+1
)
θE
[e−Vt+1θ
]
(A.34)
− σ(1 + (−kt + kt+1)φ)
(ct −
hωt
ω
)−1−σ (cχ,t − h−1+ω
t hχ,t
)+ φ
(ct −
hωt
ω
)−σkχ,t+1 = β
(E
[e−Vt+1
θ
(1 + at+1h
1−αt+1 k
−1+αt+1 α− δ + (−kt+1 + kt+2)φ
)(ct+1 −
hωt+1ω
)−σE
e−Vt+1θ
(Vχ,t+1+at+1Va,t+1+kχt+1Vk,t+1+dχ,t+1Vd,t+1
)
θ
E
[e−Vt+1θ
]2
+1
E
[e−Vt+1θ
](−σ
(1 + at+1h
1−αt+1 k
−1+αt+1 α− δ + (−kt+1 + kt+2)φ
)(ct+1 −
hωt+1
ω
)−1−σ (cχ,t+1 − h−1+ω
t+1 hχ,t+1
)
+
(ct+1 −
hωt+1
ω
)−σ (α(h1−αt+1 k
−1+αt+1 aχ,t+1 + at+1
(h−αt+1k
−1+αt+1 (1− α)hχ,t+1 + h
1−αt+1 k
−2+αt+1 (−1 + α)kχ,t+1
))
+φ(−kχt+1 + kχ,t+2 + aχ,t+1ka,t+2 + kt+1kk,t+2 + dχ,t+1kd,t+2
))))]− E
[e−Vt+1θ
(1 + at+1h
1−αt+1 k
−1+αt+1 α− δ + (−kt+1 + kt+2)φ
)(ct+1 −
hωt+1ω
)−σ (Vχ,t+1 + aχ,t+1Va,t+1 + kχ,t+1Vk,t+1 + dχ,t+1Vd,t+1
)
θE
[e−Vt+1θ
]
(A.35)
A.3 Simulation results
The results in the following table show the compensating variation in consumption τ for dif-
ferent values of risk aversion σ and the concern for robustness θ. Calculations for the benchmark
scenario are performed with Schmitt-Grohe and Uribe’s (2003) original calibration. The second
scenario, assumes financial frictions understood as a high sensitivity of domestic interest rate to
debt. Parameter characterizing the sensitivity of the sovereign risk to domestic debt stock (ψ) is
approximately 1000 times greater than the original scenario.
Table A.1: Welfare loss estimations due to aggregate uncertainty, for different combinations of riskaversion and concern for model uncertainty; no financial frictions case (consumption equivalentvariation, τ)
APPENDIX A. APPENDIX WELFARE COST OF MODEL UNCERTAINTY IN A SOE 50
Risk Aversion Model
Risk (θ =∞) Risk+MU (θ = 8) Risk+MU (θ = 1.2)
σ = 0.1 -0.01242487494 -0.0047773802 0.02940232963
σ = 0.5 -0.0104147361 -0.00031245048 0.04775196528
σ = 1 -0.00827278276 -0.00476982253 0.07859008227
σ = 1.5 -0.0062126702 0.01546052899 0.1288140583
σ = 2 -0.0041754876 0.0282842252 0.2021151683
σ = 3 -0.0001189552 0.07435342758 0.4793921136
σ = 4 0.003933465942 0.1769271301 1.092818428
σ = 5 0.007985148696 0.4081522957 2.332218983
Table A.2: Welfare loss estimations due to aggregate uncertainty, for different combination ofrisk aversion and concern for model uncertainty; with financial frictions (consumption equivalentvariation, τ)
Risk Aversion Model
Risk (θ =∞) Risk+MU (θ = 8) Risk+MU (θ = 1.2)
σ = 0.1 -0.01136340832 -0.00387704322 0.02937390047
σ = 0.5 -0.0078994127 0.001558175204 0.04598087782
σ = 1 -0.00479431469 0.006784700022 0.0723374851
σ = 1.5 -0.00213717794 0.01675790425 0.1144456391
σ = 2 0.000296700405 0.0278397621 0.1740893674
σ = 3 0.00480046866624 0.06535471154 0.3941810435
σ = 4 0.009047364944 0.1453796449 0.8741916885
σ = 5 0.01313991887 0.3218276507 1.856003205